HAL Id hal-02357327httpshalarchives-ouvertesfrhal-02357327

Submitted on 24 Nov 2019

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents whether they are pub-lished or not The documents may come fromteaching and research institutions in France orabroad or from public or private research centers

Lrsquoarchive ouverte pluridisciplinaire HAL estdestineacutee au deacutepocirct et agrave la diffusion de documentsscientifiques de niveau recherche publieacutes ou noneacutemanant des eacutetablissements drsquoenseignement et derecherche franccedilais ou eacutetrangers des laboratoirespublics ou priveacutes

LiDAR derived forest structure data improvespredictions of canopy N and P concentrations from

imaging spectroscopyMichael Ewald Raf Aerts Jonathan Lenoir Fabian Ewald Fassnacht Manuel

Nicolas Sandra Skowronek Jerome Piat Olivier Honnay Carol XimenaGarzon-Lopez Hannes Feilhauer et al

To cite this versionMichael Ewald Raf Aerts Jonathan Lenoir Fabian Ewald Fassnacht Manuel Nicolas et alLiDAR derived forest structure data improves predictions of canopy N and P concentrationsfrom imaging spectroscopy Remote Sensing of Environment Elsevier 2018 211 pp13ndash25101016jrse201803038 hal-02357327

LiDAR derived forest structure dataimproves predictions of canopy N and P

concentrations from imaging spectroscopyMichael Ewald1 Raf Aerts2 Jonathan Lenoir3 Fabian Ewald Fassnacht1 Manuel

Nicolas4 Sandra Skowronek5 Jeacuterocircme Piat4 Olivier Honnay2 Carol XimenaGarzoacuten-Loacutepez36 Hannes Feilhauer5 Ruben Van De Kerchove7 Ben Somers8 Tarek

Hattab39 Duccio Rocchini101112 and Sebastian Schmidtlein1

1Institute of Geography and Geoecology Karlsruhe Institute of Technology (KIT)Kaiserstr 12 76131 Karlsruhe Germany

2Biology Department KU Leuven Kasteelpark Arenberg 31-2435 3001 Leuven Belgium3UR ldquoEcologie et Dynamique des Systegravemes Anthropiseacutesrdquo (EDYSAN UMR 7058 CNRS)Universiteacute de Picardie Jules Verne 1 rue des Louvels 80037 Amiens Cedex 1 France4Deacutepartement Recherche et Degraveveloppement Office National des Forecircts Boulevard de

Constance 77300 Fontainebleau France5Institute of Geography FAU Erlangen-Nuremberg Wetterkreuz 15 91058 Erlangen

Germany6Ecology and Vegetation physiology group (EcoFiv) Universidad de los Andes Cr 1E

No 18ABogotaacute Colombia7VITO Flemish institute for technological research Boeretang 200 2400 Mol Belgium

8Department of Earth amp Environmental Sciences KU Leuven Celestijnenlaan 200E3001 Leuven Belgium

9Institut Franccedilais de Recherche pour lrsquoExploitation de la Mer UMR MARBEC AvenueJean Monnet CS Segravete France

10Department of Biodiversity and Molecular Ecology Research and Innovation CentreFondazione Edmund Mach Via E Mach 1 38010 San Michele allrsquoAdige TN Italy

11Center Agriculture Food Environment University of Trento Via E Mach 1 38010 SMichele allrsquoAdige (TN) Italy

12Centre for Integrative Biology University of Trento Via Sommarive 14 38123 Povo(TN) Italy

Imaging spectroscopy is a powerful tool for mapping chemical leaf traitsat the canopy level However covariance with structural canopy propertiesis hampering the ability to predict leaf biochemical traits in structurally het-erogeneous forests Here we used imaging spectroscopy data to map canopy

1

level leaf nitrogen (Nmass) and phosphorus concentrations (Pmass) of a tem-perate mixed forest By integrating predictor variables derived from airbornelaser scanning (LiDAR) capturing the biophysical complexity of the canopywe aimed at improving predictions of Nmass and Pmass We used partialleast squared regression (PLSR) models to link community weighted meansof both leaf constituents with 245 hyperspectral bands (450 - 2450 nm) and38 LiDAR-derived variables LiDAR-derived variables improved the modelrsquosexplained variances for Nmass (R2

cv 031 vs 041 RMSEcv 33 vs 30)and Pmass (R2

cv 045 vs 063 RMSEcv 153 vs 125) The predictiveperformances of Nmass models using hyperspectral bands only decreasedwith increasing structural heterogeneity included in the calibration datasetTo test the independent contribution of canopy structure we additionally fitthe models using only LiDAR-derived variables as predictors Resulting R2

cv

values ranged from 026 for Nmass to 054 for Pmass indicating considerablecovariation between these biochemical traits and forest structural propertiesNmass was negatively related to the spatial heterogeneity of canopy densitywhereas Pmass was negatively related to canopy height and to the total coverof tree canopies In the specific setting of this study the importance ofstructural variables can be attributed to the presence of two tree speciesfeaturing structural and biochemical properties different from co-occurringspecies Still existing functional linkages between structure and biochem-istry at the leaf and canopy level suggest that canopy structure used asproxy can in general support the mapping of leaf biochemistry over broadspatial extents

1 IntroductionPlant traits are important indicators of ecosystem functioning and are widely used inecological research to detect responses to environmental change (Chapin 2003 Garnieret al 2007 Kimberley et al 2014) or to quantify ecosystem services (Lamarque et al2014 Lavorel et al 2011) Biochemical traits like leaf nitrogen and phosphorus contentrespond to changing environmental conditions such as soil nutrients or climate (Di Paloand Fornara 2015 Sardans et al 2015) and are key factors related to important ecolog-ical processes including net primary production and litter deiosition (Melillo et al 1982Ollinger et al 2002 Reich 2012) Temporal trends like increasing NP ratios causedby nitrogen deposition can serve as indicators for ecosystem health and sustainability(Jonard et al 2015 Talkner et al 2015) Using leaf traits to answer questions relatedto ecosystem functioning often requires scaling from the leaf to the plant community orecosystem level (Masek et al 2015 Suding et al 2008) Due to the fact that certain leafbiochemical traits are closely linked to the reflectance signature of leaves (Kokaly et al2009) the use of imaging spectroscopy has proved to be an efficient method for scalingand the prediction of these traits across large spatial scales (Homolovaacute et al 2013) Byfar most studies relating foliage biochemistry to airborne imaging spectroscopy data

2

focused on leaf nitrogen (eg Dahlin et al 2013 Huber et al 2008 Martin and Aber1997 Wang et al 2016) But also other biochemical leaf ingredients like chlorophyllcellulose and lignin (Curran et al 1997 Schlerf et al 2010 Serrano et al 2002) andeven micronutrients like iron and copper (Asner et al 2015 Pullanagari et al 2016)have been successfully related to imaging spectroscopy data Compared to leaf nitrogenmapping of leaf phosphorus concentrations received less attention (but see Asner et al2015 Porder et al 2005 Pullanagari et al 2016)The link between leaf biochemistry and reflectance established in optical remote sens-

ing applications strongly depends on the observational level At the leaf level nitrogenconcentrations for example are directly expressed in the spectral signal For dried andground samples characteristic absorption features can be found in the shortwave in-frared (SWIR) region of the electromagnetic spectrum The absorption of radiation inthe SWIR can be attributed to nitrogen bonds in organic compounds primarily of leafproteins (Kokaly et al 2009) In fresh leaves the nitrogen concentration is addition-ally strongly related to absorption in the visible part of the spectrum (VIS) (Asner andMartin 2008) which can be attributed to the correlation between chlorophyll and leafnitrogen (Homolovaacute et al 2013 Ollinger 2011) At the canopy level spectral reflectanceis strongly influenced by canopy structure (Asner 1998 Gerard and North 1997 Rauti-ainen et al 2004) Thus the estimation of leaf traits from canopy reflectance is morecomplex due to the confounding effects of structural properties like crown morphologyleaf area index (LAI) leaf clumping or stand height (Ali et al 2016 Simic et al 2011Xiao et al 2014) Consequently variability in canopy structure can strongly influencethe accuracy of nitrogen estimations from remote sensing (Asner and Martin 2008) Onthe other hand canopy structure has been found to explain part of the relation betweenreflectance and canopy nitrogen This relation is revealed by a strong importance ofreflectance in the near infrared (NIR) for mapping canopy nitrogen reported by previ-ous studies (Martin et al 2008 Ollinger et al 2008) Reflection in the NIR region isdominated by multiple scattering between leaves of the canopy and thus very sensitiveto variation in canopy structure (Knyazikhin et al 2013 Ollinger 2011) Covariationbetween canopy structure and nitrogen was found across different types of forest ecosys-tems and hence points at the existence of a functional link between canopy structureand biochemical composition However the foundation of this functional link has notbeen fully understoodIn this study we aim at scaling leaf level measurements of mass based leaf nitrogen

(Nmass) and phosphorus content (Pmass) to the canopy scale for a temperate mixed for-est To capture the forestrsquos diversity in terms of tree species age distribution and canopystructure we propose to explicitly integrate information on forest structure derived fromairborne laser scanning (Light Detection And Ranging LiDAR) into the empirical mod-els Airborne LiDAR data can depict the 3D structure of the vegetation and has beensuccessfully used to map forest attributes like the leaf area index and standing biomass(Fassnacht et al 2014 Korhonen et al 2011 Zolkos et al 2013) The benefit of LiDAR-derived information on forest structure for mapping of canopy biochemistry has not beenassessed yet We argue that the integration of structural properties allows for a betteracquisition of leaf chemical traits in heterogeneous forests canopies We furthermore

3

expect that LiDAR data can help to understand expected covariation between canopystructural properties and biochemical leaf traits Specifically we aim at (1) improvingpredictions of Nmass and Pmass using imaging spectroscopy through the integration ofLiDAR-derived information on forest structure and (2) finding out which structuralcanopy properties correlate with Nmass and Pmass in canopies of mixed forests

2 Materials and Methods21 Study areaThe study area is the forest of Compiegravegne (northern France 49370 N 2886 E) cov-ering an area of 1442 km2 This lowland forest is located in the humid temperateclimate zone with a mean annual temperature of 103Cand mean annual precipitationof 677mm The soils cover a range from acidic nutrient-poor sandy soils to basic andhydromorphic soils (Closset-Kopp et al 2010) The forest mainly consists of even-agedmanaged stands of beech (Fagus sylvatica) oaks (Quercus robur Quercus petraea) andpine (Pinus sylvestris) growing in mono-culture as well as in mixed stands frequentlyintermingled with European hornbeam (Carpinus betulus) and ash (Fraxinius excelsior)(Chabrerie et al 2008) Stands are covering a range from early pioneer stages to morethan 200-year-old mature forests As a result of thinning activities and windthrow theforest is characterized by frequent canopy gaps which are often filled by the Americanblack cherry (Prunus serotina) an alien invasive tree species in central Europe Prunusserotina is in some parts also highly abundant in the upper canopy of earlier pioneerstages

22 Field dataField data were acquired from 50 north-facing field plots (25m times 25m) established inJuly 2014 Of those plots 44 plots were randomly selected from an initial set of 64field plots established in 2004 during a previous field study by Chabrerie et al (2008)Six additional plots were selected to include stands in earlier stages of forest successionaiming to cover the entire range of structural canopy complexity The plots coveredall main forest stand types including mixed tree species stands in different age classes(supplementary material Tab S1) In each plot we recorded the diameter at breastheight for all trees and shrubs higher than 2mIn July 2015 we sampled leaves from the most abundant tree species making up at

least 80 of the basal area in one plot This resulted in up to five sampled species perplot For each species in each plot we took three independent samples if possible fromdifferent individuals Taller trees were sampled by shooting branches using shotguns(Marlin Model 55 Goose Marlin Firearms Co Madison USA and Winchester SelectSporting II 12MWinchester Morgan USA) with Buckshot 27 ammunition (27 times 62mmpellets) aiming at single branches (Aerts et al 2017) Samples from smaller trees weretaken using a pole clipper In both cases leaves from the upper part of the crown werepreferably chosen Trees growing in canopy gaps were sampled in the center of these

4

gaps in order to collect the most sunlit leaves from these individuals For broadleavedtrees each sample consisted of 10 to 15 undamaged leaves depending on leaf size Thesamples of the only coniferous tree species P sylvestris consisted of at least 20 needlesfrom both the current and the last growing season In total we collected 328 leaf samplesfrom nine different tree species Leaves were put in sealed plastic bags and stored incooling boxes At the end of each field day samples were weighed and then dried at80C for 48 hoursBack from the field leaves were milled prior to the analysis Nmass was measured

applying the Dumas method using a vario MACRO element analyzer (Elementar Anal-ysensysteme Hanau Germany) Pmass was measured using an inductively coupledplasma-optical emission spectrometer (ICP-OES) (Varian 725ES Varian Inc Palo AltoCA USA) For each field plot we calculated community weighted mean values for Nmass

and Pmass taking the basal area of each species in the corresponding plot as the weightThe relative basal area is a good approximation for relative canopy cover of the treespecies co-occurring in a forest stand (Cade 1997 Gill et al 2000) The relative canopycover corresponds to the contribution of each species to the reflectance signal of a mixedforest canopy Although field samples were collected one year after the acquisition ofremote sensing data we consider our field data set as a solid basis for the prediction ofNmass and Pmass Previous studies indicate that in temperate tree species there are noremarkable differences in leaf chemical contents between two consecutive years (Reichet al 1991 Smith et al 2003) Furthermore Nmass in deciduous broadleaved speciestypically shows only little variation during the mid-growing season (McKown et al2013 Niinemets 2016 Reich et al 1991) and remains stable under drought conditions(Grassi et al 2005 Wilson et al 2000) The latter point is noteworthy because theearly summer of 2015 was dryer compared to the year 2014

23 Remote sensing dataWe used airborne imaging spectroscopy data (284 bands 380 nm ndash 2500 nm) acquiredby the Airborne Prism Experiment (APEX) spectrometer with a spatial resolution of3m times 3m and airborne discrete return LiDAR data with an average point density of 23points per m2 both covering the entire study area APEX data were acquired on July24 2014 (956 ndash 1125 UTC + 2h) at a flight height of 5400m by the Flemish Institute ofTechnology (VITO Mol Belgium) The data consisting of 12 flight lines were deliveredgeometrically and atmospherically corrected using the standard processing chain appliedto APEX recorded images (Sterckx et al 2016 Vreys et al 2016) Bands from both endsof the spectra and bands disturbed by water absorption were deleted prior to the analysisIn total we included 245 spectral bands between 426 nm and 2425 nm for subsequentanalyses We applied a Normalized Differenced Vegetation Index (NDVI) mask in orderto exclude values from pixels with bare soil and ground vegetation (Asner et al 2015)For this purpose we calculated NDVI values for each pixel and excluded pixels witha NDVI below 075 For all remaining pixels we applied a brightness normalization toreduce the influence of canopy shades on the spectral signal (Feilhauer et al 2010)LiDAR points were recorded in February 2014 at leaf-off conditions by Aerodata (Lille

5

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

level leaf nitrogen (Nmass) and phosphorus concentrations (Pmass) of a tem-perate mixed forest By integrating predictor variables derived from airbornelaser scanning (LiDAR) capturing the biophysical complexity of the canopywe aimed at improving predictions of Nmass and Pmass We used partialleast squared regression (PLSR) models to link community weighted meansof both leaf constituents with 245 hyperspectral bands (450 - 2450 nm) and38 LiDAR-derived variables LiDAR-derived variables improved the modelrsquosexplained variances for Nmass (R2

cv 031 vs 041 RMSEcv 33 vs 30)and Pmass (R2

cv 045 vs 063 RMSEcv 153 vs 125) The predictiveperformances of Nmass models using hyperspectral bands only decreasedwith increasing structural heterogeneity included in the calibration datasetTo test the independent contribution of canopy structure we additionally fitthe models using only LiDAR-derived variables as predictors Resulting R2

cv

values ranged from 026 for Nmass to 054 for Pmass indicating considerablecovariation between these biochemical traits and forest structural propertiesNmass was negatively related to the spatial heterogeneity of canopy densitywhereas Pmass was negatively related to canopy height and to the total coverof tree canopies In the specific setting of this study the importance ofstructural variables can be attributed to the presence of two tree speciesfeaturing structural and biochemical properties different from co-occurringspecies Still existing functional linkages between structure and biochem-istry at the leaf and canopy level suggest that canopy structure used asproxy can in general support the mapping of leaf biochemistry over broadspatial extents

1 IntroductionPlant traits are important indicators of ecosystem functioning and are widely used inecological research to detect responses to environmental change (Chapin 2003 Garnieret al 2007 Kimberley et al 2014) or to quantify ecosystem services (Lamarque et al2014 Lavorel et al 2011) Biochemical traits like leaf nitrogen and phosphorus contentrespond to changing environmental conditions such as soil nutrients or climate (Di Paloand Fornara 2015 Sardans et al 2015) and are key factors related to important ecolog-ical processes including net primary production and litter deiosition (Melillo et al 1982Ollinger et al 2002 Reich 2012) Temporal trends like increasing NP ratios causedby nitrogen deposition can serve as indicators for ecosystem health and sustainability(Jonard et al 2015 Talkner et al 2015) Using leaf traits to answer questions relatedto ecosystem functioning often requires scaling from the leaf to the plant community orecosystem level (Masek et al 2015 Suding et al 2008) Due to the fact that certain leafbiochemical traits are closely linked to the reflectance signature of leaves (Kokaly et al2009) the use of imaging spectroscopy has proved to be an efficient method for scalingand the prediction of these traits across large spatial scales (Homolovaacute et al 2013) Byfar most studies relating foliage biochemistry to airborne imaging spectroscopy data

2

focused on leaf nitrogen (eg Dahlin et al 2013 Huber et al 2008 Martin and Aber1997 Wang et al 2016) But also other biochemical leaf ingredients like chlorophyllcellulose and lignin (Curran et al 1997 Schlerf et al 2010 Serrano et al 2002) andeven micronutrients like iron and copper (Asner et al 2015 Pullanagari et al 2016)have been successfully related to imaging spectroscopy data Compared to leaf nitrogenmapping of leaf phosphorus concentrations received less attention (but see Asner et al2015 Porder et al 2005 Pullanagari et al 2016)The link between leaf biochemistry and reflectance established in optical remote sens-

ing applications strongly depends on the observational level At the leaf level nitrogenconcentrations for example are directly expressed in the spectral signal For dried andground samples characteristic absorption features can be found in the shortwave in-frared (SWIR) region of the electromagnetic spectrum The absorption of radiation inthe SWIR can be attributed to nitrogen bonds in organic compounds primarily of leafproteins (Kokaly et al 2009) In fresh leaves the nitrogen concentration is addition-ally strongly related to absorption in the visible part of the spectrum (VIS) (Asner andMartin 2008) which can be attributed to the correlation between chlorophyll and leafnitrogen (Homolovaacute et al 2013 Ollinger 2011) At the canopy level spectral reflectanceis strongly influenced by canopy structure (Asner 1998 Gerard and North 1997 Rauti-ainen et al 2004) Thus the estimation of leaf traits from canopy reflectance is morecomplex due to the confounding effects of structural properties like crown morphologyleaf area index (LAI) leaf clumping or stand height (Ali et al 2016 Simic et al 2011Xiao et al 2014) Consequently variability in canopy structure can strongly influencethe accuracy of nitrogen estimations from remote sensing (Asner and Martin 2008) Onthe other hand canopy structure has been found to explain part of the relation betweenreflectance and canopy nitrogen This relation is revealed by a strong importance ofreflectance in the near infrared (NIR) for mapping canopy nitrogen reported by previ-ous studies (Martin et al 2008 Ollinger et al 2008) Reflection in the NIR region isdominated by multiple scattering between leaves of the canopy and thus very sensitiveto variation in canopy structure (Knyazikhin et al 2013 Ollinger 2011) Covariationbetween canopy structure and nitrogen was found across different types of forest ecosys-tems and hence points at the existence of a functional link between canopy structureand biochemical composition However the foundation of this functional link has notbeen fully understoodIn this study we aim at scaling leaf level measurements of mass based leaf nitrogen

(Nmass) and phosphorus content (Pmass) to the canopy scale for a temperate mixed for-est To capture the forestrsquos diversity in terms of tree species age distribution and canopystructure we propose to explicitly integrate information on forest structure derived fromairborne laser scanning (Light Detection And Ranging LiDAR) into the empirical mod-els Airborne LiDAR data can depict the 3D structure of the vegetation and has beensuccessfully used to map forest attributes like the leaf area index and standing biomass(Fassnacht et al 2014 Korhonen et al 2011 Zolkos et al 2013) The benefit of LiDAR-derived information on forest structure for mapping of canopy biochemistry has not beenassessed yet We argue that the integration of structural properties allows for a betteracquisition of leaf chemical traits in heterogeneous forests canopies We furthermore

3

expect that LiDAR data can help to understand expected covariation between canopystructural properties and biochemical leaf traits Specifically we aim at (1) improvingpredictions of Nmass and Pmass using imaging spectroscopy through the integration ofLiDAR-derived information on forest structure and (2) finding out which structuralcanopy properties correlate with Nmass and Pmass in canopies of mixed forests

2 Materials and Methods21 Study areaThe study area is the forest of Compiegravegne (northern France 49370 N 2886 E) cov-ering an area of 1442 km2 This lowland forest is located in the humid temperateclimate zone with a mean annual temperature of 103Cand mean annual precipitationof 677mm The soils cover a range from acidic nutrient-poor sandy soils to basic andhydromorphic soils (Closset-Kopp et al 2010) The forest mainly consists of even-agedmanaged stands of beech (Fagus sylvatica) oaks (Quercus robur Quercus petraea) andpine (Pinus sylvestris) growing in mono-culture as well as in mixed stands frequentlyintermingled with European hornbeam (Carpinus betulus) and ash (Fraxinius excelsior)(Chabrerie et al 2008) Stands are covering a range from early pioneer stages to morethan 200-year-old mature forests As a result of thinning activities and windthrow theforest is characterized by frequent canopy gaps which are often filled by the Americanblack cherry (Prunus serotina) an alien invasive tree species in central Europe Prunusserotina is in some parts also highly abundant in the upper canopy of earlier pioneerstages

22 Field dataField data were acquired from 50 north-facing field plots (25m times 25m) established inJuly 2014 Of those plots 44 plots were randomly selected from an initial set of 64field plots established in 2004 during a previous field study by Chabrerie et al (2008)Six additional plots were selected to include stands in earlier stages of forest successionaiming to cover the entire range of structural canopy complexity The plots coveredall main forest stand types including mixed tree species stands in different age classes(supplementary material Tab S1) In each plot we recorded the diameter at breastheight for all trees and shrubs higher than 2mIn July 2015 we sampled leaves from the most abundant tree species making up at

least 80 of the basal area in one plot This resulted in up to five sampled species perplot For each species in each plot we took three independent samples if possible fromdifferent individuals Taller trees were sampled by shooting branches using shotguns(Marlin Model 55 Goose Marlin Firearms Co Madison USA and Winchester SelectSporting II 12MWinchester Morgan USA) with Buckshot 27 ammunition (27 times 62mmpellets) aiming at single branches (Aerts et al 2017) Samples from smaller trees weretaken using a pole clipper In both cases leaves from the upper part of the crown werepreferably chosen Trees growing in canopy gaps were sampled in the center of these

4

gaps in order to collect the most sunlit leaves from these individuals For broadleavedtrees each sample consisted of 10 to 15 undamaged leaves depending on leaf size Thesamples of the only coniferous tree species P sylvestris consisted of at least 20 needlesfrom both the current and the last growing season In total we collected 328 leaf samplesfrom nine different tree species Leaves were put in sealed plastic bags and stored incooling boxes At the end of each field day samples were weighed and then dried at80C for 48 hoursBack from the field leaves were milled prior to the analysis Nmass was measured

applying the Dumas method using a vario MACRO element analyzer (Elementar Anal-ysensysteme Hanau Germany) Pmass was measured using an inductively coupledplasma-optical emission spectrometer (ICP-OES) (Varian 725ES Varian Inc Palo AltoCA USA) For each field plot we calculated community weighted mean values for Nmass

and Pmass taking the basal area of each species in the corresponding plot as the weightThe relative basal area is a good approximation for relative canopy cover of the treespecies co-occurring in a forest stand (Cade 1997 Gill et al 2000) The relative canopycover corresponds to the contribution of each species to the reflectance signal of a mixedforest canopy Although field samples were collected one year after the acquisition ofremote sensing data we consider our field data set as a solid basis for the prediction ofNmass and Pmass Previous studies indicate that in temperate tree species there are noremarkable differences in leaf chemical contents between two consecutive years (Reichet al 1991 Smith et al 2003) Furthermore Nmass in deciduous broadleaved speciestypically shows only little variation during the mid-growing season (McKown et al2013 Niinemets 2016 Reich et al 1991) and remains stable under drought conditions(Grassi et al 2005 Wilson et al 2000) The latter point is noteworthy because theearly summer of 2015 was dryer compared to the year 2014

23 Remote sensing dataWe used airborne imaging spectroscopy data (284 bands 380 nm ndash 2500 nm) acquiredby the Airborne Prism Experiment (APEX) spectrometer with a spatial resolution of3m times 3m and airborne discrete return LiDAR data with an average point density of 23points per m2 both covering the entire study area APEX data were acquired on July24 2014 (956 ndash 1125 UTC + 2h) at a flight height of 5400m by the Flemish Institute ofTechnology (VITO Mol Belgium) The data consisting of 12 flight lines were deliveredgeometrically and atmospherically corrected using the standard processing chain appliedto APEX recorded images (Sterckx et al 2016 Vreys et al 2016) Bands from both endsof the spectra and bands disturbed by water absorption were deleted prior to the analysisIn total we included 245 spectral bands between 426 nm and 2425 nm for subsequentanalyses We applied a Normalized Differenced Vegetation Index (NDVI) mask in orderto exclude values from pixels with bare soil and ground vegetation (Asner et al 2015)For this purpose we calculated NDVI values for each pixel and excluded pixels witha NDVI below 075 For all remaining pixels we applied a brightness normalization toreduce the influence of canopy shades on the spectral signal (Feilhauer et al 2010)LiDAR points were recorded in February 2014 at leaf-off conditions by Aerodata (Lille

5

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

focused on leaf nitrogen (eg Dahlin et al 2013 Huber et al 2008 Martin and Aber1997 Wang et al 2016) But also other biochemical leaf ingredients like chlorophyllcellulose and lignin (Curran et al 1997 Schlerf et al 2010 Serrano et al 2002) andeven micronutrients like iron and copper (Asner et al 2015 Pullanagari et al 2016)have been successfully related to imaging spectroscopy data Compared to leaf nitrogenmapping of leaf phosphorus concentrations received less attention (but see Asner et al2015 Porder et al 2005 Pullanagari et al 2016)The link between leaf biochemistry and reflectance established in optical remote sens-

ing applications strongly depends on the observational level At the leaf level nitrogenconcentrations for example are directly expressed in the spectral signal For dried andground samples characteristic absorption features can be found in the shortwave in-frared (SWIR) region of the electromagnetic spectrum The absorption of radiation inthe SWIR can be attributed to nitrogen bonds in organic compounds primarily of leafproteins (Kokaly et al 2009) In fresh leaves the nitrogen concentration is addition-ally strongly related to absorption in the visible part of the spectrum (VIS) (Asner andMartin 2008) which can be attributed to the correlation between chlorophyll and leafnitrogen (Homolovaacute et al 2013 Ollinger 2011) At the canopy level spectral reflectanceis strongly influenced by canopy structure (Asner 1998 Gerard and North 1997 Rauti-ainen et al 2004) Thus the estimation of leaf traits from canopy reflectance is morecomplex due to the confounding effects of structural properties like crown morphologyleaf area index (LAI) leaf clumping or stand height (Ali et al 2016 Simic et al 2011Xiao et al 2014) Consequently variability in canopy structure can strongly influencethe accuracy of nitrogen estimations from remote sensing (Asner and Martin 2008) Onthe other hand canopy structure has been found to explain part of the relation betweenreflectance and canopy nitrogen This relation is revealed by a strong importance ofreflectance in the near infrared (NIR) for mapping canopy nitrogen reported by previ-ous studies (Martin et al 2008 Ollinger et al 2008) Reflection in the NIR region isdominated by multiple scattering between leaves of the canopy and thus very sensitiveto variation in canopy structure (Knyazikhin et al 2013 Ollinger 2011) Covariationbetween canopy structure and nitrogen was found across different types of forest ecosys-tems and hence points at the existence of a functional link between canopy structureand biochemical composition However the foundation of this functional link has notbeen fully understoodIn this study we aim at scaling leaf level measurements of mass based leaf nitrogen

(Nmass) and phosphorus content (Pmass) to the canopy scale for a temperate mixed for-est To capture the forestrsquos diversity in terms of tree species age distribution and canopystructure we propose to explicitly integrate information on forest structure derived fromairborne laser scanning (Light Detection And Ranging LiDAR) into the empirical mod-els Airborne LiDAR data can depict the 3D structure of the vegetation and has beensuccessfully used to map forest attributes like the leaf area index and standing biomass(Fassnacht et al 2014 Korhonen et al 2011 Zolkos et al 2013) The benefit of LiDAR-derived information on forest structure for mapping of canopy biochemistry has not beenassessed yet We argue that the integration of structural properties allows for a betteracquisition of leaf chemical traits in heterogeneous forests canopies We furthermore

3

expect that LiDAR data can help to understand expected covariation between canopystructural properties and biochemical leaf traits Specifically we aim at (1) improvingpredictions of Nmass and Pmass using imaging spectroscopy through the integration ofLiDAR-derived information on forest structure and (2) finding out which structuralcanopy properties correlate with Nmass and Pmass in canopies of mixed forests

2 Materials and Methods21 Study areaThe study area is the forest of Compiegravegne (northern France 49370 N 2886 E) cov-ering an area of 1442 km2 This lowland forest is located in the humid temperateclimate zone with a mean annual temperature of 103Cand mean annual precipitationof 677mm The soils cover a range from acidic nutrient-poor sandy soils to basic andhydromorphic soils (Closset-Kopp et al 2010) The forest mainly consists of even-agedmanaged stands of beech (Fagus sylvatica) oaks (Quercus robur Quercus petraea) andpine (Pinus sylvestris) growing in mono-culture as well as in mixed stands frequentlyintermingled with European hornbeam (Carpinus betulus) and ash (Fraxinius excelsior)(Chabrerie et al 2008) Stands are covering a range from early pioneer stages to morethan 200-year-old mature forests As a result of thinning activities and windthrow theforest is characterized by frequent canopy gaps which are often filled by the Americanblack cherry (Prunus serotina) an alien invasive tree species in central Europe Prunusserotina is in some parts also highly abundant in the upper canopy of earlier pioneerstages

22 Field dataField data were acquired from 50 north-facing field plots (25m times 25m) established inJuly 2014 Of those plots 44 plots were randomly selected from an initial set of 64field plots established in 2004 during a previous field study by Chabrerie et al (2008)Six additional plots were selected to include stands in earlier stages of forest successionaiming to cover the entire range of structural canopy complexity The plots coveredall main forest stand types including mixed tree species stands in different age classes(supplementary material Tab S1) In each plot we recorded the diameter at breastheight for all trees and shrubs higher than 2mIn July 2015 we sampled leaves from the most abundant tree species making up at

least 80 of the basal area in one plot This resulted in up to five sampled species perplot For each species in each plot we took three independent samples if possible fromdifferent individuals Taller trees were sampled by shooting branches using shotguns(Marlin Model 55 Goose Marlin Firearms Co Madison USA and Winchester SelectSporting II 12MWinchester Morgan USA) with Buckshot 27 ammunition (27 times 62mmpellets) aiming at single branches (Aerts et al 2017) Samples from smaller trees weretaken using a pole clipper In both cases leaves from the upper part of the crown werepreferably chosen Trees growing in canopy gaps were sampled in the center of these

4

gaps in order to collect the most sunlit leaves from these individuals For broadleavedtrees each sample consisted of 10 to 15 undamaged leaves depending on leaf size Thesamples of the only coniferous tree species P sylvestris consisted of at least 20 needlesfrom both the current and the last growing season In total we collected 328 leaf samplesfrom nine different tree species Leaves were put in sealed plastic bags and stored incooling boxes At the end of each field day samples were weighed and then dried at80C for 48 hoursBack from the field leaves were milled prior to the analysis Nmass was measured

applying the Dumas method using a vario MACRO element analyzer (Elementar Anal-ysensysteme Hanau Germany) Pmass was measured using an inductively coupledplasma-optical emission spectrometer (ICP-OES) (Varian 725ES Varian Inc Palo AltoCA USA) For each field plot we calculated community weighted mean values for Nmass

and Pmass taking the basal area of each species in the corresponding plot as the weightThe relative basal area is a good approximation for relative canopy cover of the treespecies co-occurring in a forest stand (Cade 1997 Gill et al 2000) The relative canopycover corresponds to the contribution of each species to the reflectance signal of a mixedforest canopy Although field samples were collected one year after the acquisition ofremote sensing data we consider our field data set as a solid basis for the prediction ofNmass and Pmass Previous studies indicate that in temperate tree species there are noremarkable differences in leaf chemical contents between two consecutive years (Reichet al 1991 Smith et al 2003) Furthermore Nmass in deciduous broadleaved speciestypically shows only little variation during the mid-growing season (McKown et al2013 Niinemets 2016 Reich et al 1991) and remains stable under drought conditions(Grassi et al 2005 Wilson et al 2000) The latter point is noteworthy because theearly summer of 2015 was dryer compared to the year 2014

23 Remote sensing dataWe used airborne imaging spectroscopy data (284 bands 380 nm ndash 2500 nm) acquiredby the Airborne Prism Experiment (APEX) spectrometer with a spatial resolution of3m times 3m and airborne discrete return LiDAR data with an average point density of 23points per m2 both covering the entire study area APEX data were acquired on July24 2014 (956 ndash 1125 UTC + 2h) at a flight height of 5400m by the Flemish Institute ofTechnology (VITO Mol Belgium) The data consisting of 12 flight lines were deliveredgeometrically and atmospherically corrected using the standard processing chain appliedto APEX recorded images (Sterckx et al 2016 Vreys et al 2016) Bands from both endsof the spectra and bands disturbed by water absorption were deleted prior to the analysisIn total we included 245 spectral bands between 426 nm and 2425 nm for subsequentanalyses We applied a Normalized Differenced Vegetation Index (NDVI) mask in orderto exclude values from pixels with bare soil and ground vegetation (Asner et al 2015)For this purpose we calculated NDVI values for each pixel and excluded pixels witha NDVI below 075 For all remaining pixels we applied a brightness normalization toreduce the influence of canopy shades on the spectral signal (Feilhauer et al 2010)LiDAR points were recorded in February 2014 at leaf-off conditions by Aerodata (Lille

5

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

expect that LiDAR data can help to understand expected covariation between canopystructural properties and biochemical leaf traits Specifically we aim at (1) improvingpredictions of Nmass and Pmass using imaging spectroscopy through the integration ofLiDAR-derived information on forest structure and (2) finding out which structuralcanopy properties correlate with Nmass and Pmass in canopies of mixed forests

2 Materials and Methods21 Study areaThe study area is the forest of Compiegravegne (northern France 49370 N 2886 E) cov-ering an area of 1442 km2 This lowland forest is located in the humid temperateclimate zone with a mean annual temperature of 103Cand mean annual precipitationof 677mm The soils cover a range from acidic nutrient-poor sandy soils to basic andhydromorphic soils (Closset-Kopp et al 2010) The forest mainly consists of even-agedmanaged stands of beech (Fagus sylvatica) oaks (Quercus robur Quercus petraea) andpine (Pinus sylvestris) growing in mono-culture as well as in mixed stands frequentlyintermingled with European hornbeam (Carpinus betulus) and ash (Fraxinius excelsior)(Chabrerie et al 2008) Stands are covering a range from early pioneer stages to morethan 200-year-old mature forests As a result of thinning activities and windthrow theforest is characterized by frequent canopy gaps which are often filled by the Americanblack cherry (Prunus serotina) an alien invasive tree species in central Europe Prunusserotina is in some parts also highly abundant in the upper canopy of earlier pioneerstages

22 Field dataField data were acquired from 50 north-facing field plots (25m times 25m) established inJuly 2014 Of those plots 44 plots were randomly selected from an initial set of 64field plots established in 2004 during a previous field study by Chabrerie et al (2008)Six additional plots were selected to include stands in earlier stages of forest successionaiming to cover the entire range of structural canopy complexity The plots coveredall main forest stand types including mixed tree species stands in different age classes(supplementary material Tab S1) In each plot we recorded the diameter at breastheight for all trees and shrubs higher than 2mIn July 2015 we sampled leaves from the most abundant tree species making up at

least 80 of the basal area in one plot This resulted in up to five sampled species perplot For each species in each plot we took three independent samples if possible fromdifferent individuals Taller trees were sampled by shooting branches using shotguns(Marlin Model 55 Goose Marlin Firearms Co Madison USA and Winchester SelectSporting II 12MWinchester Morgan USA) with Buckshot 27 ammunition (27 times 62mmpellets) aiming at single branches (Aerts et al 2017) Samples from smaller trees weretaken using a pole clipper In both cases leaves from the upper part of the crown werepreferably chosen Trees growing in canopy gaps were sampled in the center of these

4

gaps in order to collect the most sunlit leaves from these individuals For broadleavedtrees each sample consisted of 10 to 15 undamaged leaves depending on leaf size Thesamples of the only coniferous tree species P sylvestris consisted of at least 20 needlesfrom both the current and the last growing season In total we collected 328 leaf samplesfrom nine different tree species Leaves were put in sealed plastic bags and stored incooling boxes At the end of each field day samples were weighed and then dried at80C for 48 hoursBack from the field leaves were milled prior to the analysis Nmass was measured

applying the Dumas method using a vario MACRO element analyzer (Elementar Anal-ysensysteme Hanau Germany) Pmass was measured using an inductively coupledplasma-optical emission spectrometer (ICP-OES) (Varian 725ES Varian Inc Palo AltoCA USA) For each field plot we calculated community weighted mean values for Nmass

and Pmass taking the basal area of each species in the corresponding plot as the weightThe relative basal area is a good approximation for relative canopy cover of the treespecies co-occurring in a forest stand (Cade 1997 Gill et al 2000) The relative canopycover corresponds to the contribution of each species to the reflectance signal of a mixedforest canopy Although field samples were collected one year after the acquisition ofremote sensing data we consider our field data set as a solid basis for the prediction ofNmass and Pmass Previous studies indicate that in temperate tree species there are noremarkable differences in leaf chemical contents between two consecutive years (Reichet al 1991 Smith et al 2003) Furthermore Nmass in deciduous broadleaved speciestypically shows only little variation during the mid-growing season (McKown et al2013 Niinemets 2016 Reich et al 1991) and remains stable under drought conditions(Grassi et al 2005 Wilson et al 2000) The latter point is noteworthy because theearly summer of 2015 was dryer compared to the year 2014

23 Remote sensing dataWe used airborne imaging spectroscopy data (284 bands 380 nm ndash 2500 nm) acquiredby the Airborne Prism Experiment (APEX) spectrometer with a spatial resolution of3m times 3m and airborne discrete return LiDAR data with an average point density of 23points per m2 both covering the entire study area APEX data were acquired on July24 2014 (956 ndash 1125 UTC + 2h) at a flight height of 5400m by the Flemish Institute ofTechnology (VITO Mol Belgium) The data consisting of 12 flight lines were deliveredgeometrically and atmospherically corrected using the standard processing chain appliedto APEX recorded images (Sterckx et al 2016 Vreys et al 2016) Bands from both endsof the spectra and bands disturbed by water absorption were deleted prior to the analysisIn total we included 245 spectral bands between 426 nm and 2425 nm for subsequentanalyses We applied a Normalized Differenced Vegetation Index (NDVI) mask in orderto exclude values from pixels with bare soil and ground vegetation (Asner et al 2015)For this purpose we calculated NDVI values for each pixel and excluded pixels witha NDVI below 075 For all remaining pixels we applied a brightness normalization toreduce the influence of canopy shades on the spectral signal (Feilhauer et al 2010)LiDAR points were recorded in February 2014 at leaf-off conditions by Aerodata (Lille

5

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

gaps in order to collect the most sunlit leaves from these individuals For broadleavedtrees each sample consisted of 10 to 15 undamaged leaves depending on leaf size Thesamples of the only coniferous tree species P sylvestris consisted of at least 20 needlesfrom both the current and the last growing season In total we collected 328 leaf samplesfrom nine different tree species Leaves were put in sealed plastic bags and stored incooling boxes At the end of each field day samples were weighed and then dried at80C for 48 hoursBack from the field leaves were milled prior to the analysis Nmass was measured

applying the Dumas method using a vario MACRO element analyzer (Elementar Anal-ysensysteme Hanau Germany) Pmass was measured using an inductively coupledplasma-optical emission spectrometer (ICP-OES) (Varian 725ES Varian Inc Palo AltoCA USA) For each field plot we calculated community weighted mean values for Nmass

and Pmass taking the basal area of each species in the corresponding plot as the weightThe relative basal area is a good approximation for relative canopy cover of the treespecies co-occurring in a forest stand (Cade 1997 Gill et al 2000) The relative canopycover corresponds to the contribution of each species to the reflectance signal of a mixedforest canopy Although field samples were collected one year after the acquisition ofremote sensing data we consider our field data set as a solid basis for the prediction ofNmass and Pmass Previous studies indicate that in temperate tree species there are noremarkable differences in leaf chemical contents between two consecutive years (Reichet al 1991 Smith et al 2003) Furthermore Nmass in deciduous broadleaved speciestypically shows only little variation during the mid-growing season (McKown et al2013 Niinemets 2016 Reich et al 1991) and remains stable under drought conditions(Grassi et al 2005 Wilson et al 2000) The latter point is noteworthy because theearly summer of 2015 was dryer compared to the year 2014

23 Remote sensing dataWe used airborne imaging spectroscopy data (284 bands 380 nm ndash 2500 nm) acquiredby the Airborne Prism Experiment (APEX) spectrometer with a spatial resolution of3m times 3m and airborne discrete return LiDAR data with an average point density of 23points per m2 both covering the entire study area APEX data were acquired on July24 2014 (956 ndash 1125 UTC + 2h) at a flight height of 5400m by the Flemish Institute ofTechnology (VITO Mol Belgium) The data consisting of 12 flight lines were deliveredgeometrically and atmospherically corrected using the standard processing chain appliedto APEX recorded images (Sterckx et al 2016 Vreys et al 2016) Bands from both endsof the spectra and bands disturbed by water absorption were deleted prior to the analysisIn total we included 245 spectral bands between 426 nm and 2425 nm for subsequentanalyses We applied a Normalized Differenced Vegetation Index (NDVI) mask in orderto exclude values from pixels with bare soil and ground vegetation (Asner et al 2015)For this purpose we calculated NDVI values for each pixel and excluded pixels witha NDVI below 075 For all remaining pixels we applied a brightness normalization toreduce the influence of canopy shades on the spectral signal (Feilhauer et al 2010)LiDAR points were recorded in February 2014 at leaf-off conditions by Aerodata (Lille

5

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

France) using a Riegl LMS-680i with a maximum scan angle of 30and a lateral overlapof neighboring flight lines of 65 Average flight height during LiDAR data acquisitionwas 530m resulting in a beam diameter of about 0265m The LiDAR data were de-livered including a classification of ground and vegetation returns and a digital terrainmodel (DTM) Height values of LiDAR points were normalized by subtracting values ofthe underlying DTM Vegetation returns were then aggregated into a grid with a cell sizeof 3m times 3m taking the grid matrix of the imaging spectroscopy data as reference Foreach pixel we calculated 19 different LiDAR-derived variables based on point statisticsresulting in 19 raster layers Calculated LiDAR-derived variables included basic sum-mary statistics (eg maximum height) based on the height values of LiDAR points ineach grid cell and inverse penetration ratios representing the fractional vegetation coverwithin given height thresholds (Tab 1) (Ewald et al 2014) Penetration ratios werecalculated using the following formula

vch12 = (nh2 minus nh1)nh2 (1)

where vch12 is representing the vegetation cover within the height thresholds h1 and h2(h1 lt h2) within one grid cell nh1 and nh2 represent the sum of all LiDAR points belowthe given height thresholds h1 and h2 respectively

Table 1 Variables calculated from LiDAR point clouds in 3m times 3m resolution For the use in partialleast squares regression models variables were aggregated into a grid with a cell size of 24m times 24mby calculating mean and standard deviation

LiDAR Metric Abbreviation DescriptionMinimum min_h_mean min_h_sd Basic statisticsMaximum max_h_mean max_h_sd based on theMean mean_h_mean mean_h_sd height values ofStandard deviation sd_h_mean sd_h_sd vegetation LiDARVariance var_h_mean var_h_sd pointsCoefficient of variation cov_h_mean cov_h_sd10th percentile perc10_h_mean perc10_h_sd25th percentile perc25_h_mean perc25_h_sd50th percentile perc50_h_mean perc50_h_sd75th percentile perc75_h_mean perc75_h_sd90th percentile perc90_h_mean perc90_h_sdFractional cover 05m ndash 2m fcover_05_2_mean fcover_05_2_sd Inverse penetrationFractional cover 05m ndash 60m fcover_05_60_mean fcover_05_60_sd ratios representingFractional cover 2m ndash 6m fcover_2_6_mean fcover_2_6_sd an estimate forFractional cover 2m ndash 60m fcover_2_60_mean fcover_2_60_sd fractional cover ofFractional cover 6m ndash 10m fcover_6_10_mean fcover_6_10_sd the vegetationFractional cover 6m ndash 60m fcover_6_60_mean fcover_6_60_sd within given heightFractional cover 10m ndash 20m fcover_10_20_mean fcover_10_20_sd thresholdsFractional cover 20m ndash 60m fcover_20_60_mean fcover_20_60_sd

From both imaging spectroscopy and LiDAR raster layers we extracted values fromall pixels overlapping with the 50 field plots to be used as input to the statistical modelsFor each plot we calculated the weighted mean values of 245 hyperspectral bands and

6

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

19 LiDAR-variables (Tab 1) from the extracted cell values using the percent overlapof each cell with the plot area as weight Similarly we calculated the weighted standarddeviation for LiDAR-derived variables which represent a measure of spatial heterogeneityof these variablesFor prediction we aggregated the pixels of the imaging spectroscopy and LiDAR raster

layers to a grid with a pixel size of 24m times 24m calculating the mean and the standarddeviation (for LiDAR-derived variables only) of all aggregated cells This finally resultedin a dataset containing 245 spectral bands and 38 LiDAR-derived variables (mean andstandard deviation)

24 Model calibration and validationFor both response variables Nmass and Pmass we built predictive models using theextracted values from the raster layers at plot locations as predictors We calculatedpartial least squares regression (PLSR) models with a step-wise backward model selectionprocedure implemented in the R package autopls (R Core Team 2016 Schmidtlein et al2012) The number of latent variables was chosen based on the lowest root mean squarederror (RMSE) in leave-one-out cross-validation Before model calibration predictors werenormalized dividing each predictor variable by its standard deviationTo test the benefit of LiDAR-derived data for the prediction of community weighted

means of Nmass and Pmass at the canopy level we fit two sets of models for each re-sponse variable one incorporating the hyperspectral bands only and a second one usinga combination of hyperspectral bands and LiDAR-derived variables as predictors Totest the independent contribution of LiDAR data on the predictions we additionally fita third set of models for both Nmass and Pmass including only LiDAR-derived variablesas predictors Nmass values were natural log transformed prior to the model calculations

The model calculations and predictions were embedded in a resampling procedurewith 200 permutations in order to reduce the bias in model predictions yielding to abetter comparison between the three sets of models In each permutation a subsampleof 40 out of the 50 field plots was randomly drawn without replacement and used formodel calibration and validation Each model was used to generate a prediction mapwith a grid size of 24m times 24m resulting in 200 prediction maps for each responsevariable and each of the three predictor combinations used respectively From thesemaps we calculated a median prediction map and the associated coefficient of variation(CV) representing the spatial uncertainty of model predictions (Singh et al 2015)For the assessment of the predictive performance of the models we calculated the

mean Pearson r-squared as well as the absolute and normalized root mean squared error(RMSE) between predicted and observed values of each data subset The same perfor-mance measures were calculated for each data subset in leave-one-out cross-validationdata For Nmass r-squared values and RMSE were calculated based on the log-transformeddataset The normalized RMSE was calculated by dividing the RMSE by the mean valuein the response dataset r-squared and RMSE values were used to compare the perfor-mances of models using only hyperspectral bands or a combination of hyperspectralbands and LiDAR-derived variables as predictors for Nmass and Pmass respectively

7

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

Model performance is affected by the number of variables included in the case of aPLSR the number of latent variables To check for such an effect we grouped the cor-responding models according to the number of latent variables included and comparedthe r-squared values for each group separately (supplementary material Fig S1)

3 ResultsField plots were located in forest stands with heights ranging from 3 to 40m and LAIvalues ranging from 17 to 59 (supplementary material Tab S2) Plot-wise communityweighted mean values for Nmass and Pmass ranged from 138 to 254 gmiddotkgminus1 and from082 to 193 gmiddotkgminus1 respectively Nmass of P serotina and P sylvestris were significantlydifferent from all other species (supplementary material Fig S2 and Tab S3) Con-trary we observed no differences in measured Nmass between F sylvatica Q robur andC betulus Pmass differed significantly between all species except between C betulus andQ robur (supplementary material Fig S2) Models combining structural vegetationattributes derived from airborne LiDAR with imaging spectroscopy improved predic-tions of community weighted mean values for Nmass and Pmass compared to modelsusing imaging spectroscopy data solely (Tab 2 Fig 2) In the combined Nmass modelshyperspectral bands had a significantly higher contribution (p lt 0001) to the varianceexplained compared to LiDAR-derived variables (Fig 1) By contrast in Pmass modelsLiDAR-derived variables showed a significantly higher contribution (p lt 0001) Withrespect to the selected spectral bands we observed only marginal differences betweenmodels including LiDAR-derived variables and models not including them (Figs 3 45 6)

Table 2 Results of PLSR models for Nmass and Pmass from 200 bootstraps Predictors usedpredictor variables being either hyperspectral bands (HS) or LiDAR-derived variables LV meannumber of latent variables Var mean number of selected predictor variables R2

cal mean coefficientof determination in calibration R2

cv mean coefficient of determination in validation RMSEcalaverage root mean squared error in calibration RMSEcv average root mean squared error inleave-one-out cross-validation

Response Predictors LV Var R2cal R2

cv RMSEcal RMSEcv RMSEcal RMSEcv[] []

Nmasslowast HS 58 98 047 031 009 009 29 33plusmn 010 plusmn 014 plusmn 001 plusmn 001

HS amp LiDAR 57 43 055 041 008 009 27 30plusmn 012 plusmn 016 plusmn 001 plusmn 001

LiDAR 35 8 039 026 009 010 31 34plusmn 008 plusmn 009 plusmn 001 plusmn 001

Pmass HS 63 42 059 045 015 018 131 153plusmn 015 plusmn 016 plusmn 002 plusmn 002

HS amp LiDAR 69 38 073 063 013 014 108 125plusmn 008 plusmn 010 plusmn 002 plusmn 002

LiDAR 37 9 062 054 015 017 126 140plusmn 008 plusmn 010 plusmn 001 plusmn 001

lowastnatural log-transformed

8

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

000

025

050

075

100

HS LiDAR

VIP

pro

port

ion

Nmass

HS LiDAR

Pmass

Figure 1 Relative contribution of hyperspectral bands (HS) and LiDAR variables to the varianceexplained in PLSR models for Nmass and Pmass expressed as proportion of the total VIP (VariableImportance in Projection) score

For Nmass the average R2cv values resulting from leave-one-out cross-validation for each

bootstrap model increased from 031 to 041 whereas the mean relative RSME decreasedonly moderately (see Tab 2) when adding LiDAR-derived variables Models fittedby LiDAR-derived predictors solely resulted in a mean R2

val value of 025 The mostimportant LiDAR-derived variables in the models predicting of Nmass were accordingto VIP values related to the horizontal variation of canopy cover (fcover_05_60_sdfcover_2_6_sd fcover_6_10_sd fcover_6_60_sd) (Figs 7 8) The most importantspectral bands were located in the VIS and the SWIR between 2000 and 2400 nm irre-spective of whether only imaging spectroscopy or a combination of imaging spectroscopyand LiDAR data was used (Fig 3)For Pmass average R2

cv values resulting from leave-one-out cross-validation for eachbootstrap model increased from 045 to 063 and the mean relative RSME decreasedfrom 153 to 125 (see Tab 2) when LiDAR-derived predictors were included Modelsfitted by LiDAR-derived predictors solely resulted in a mean R2

cv value of 054 Re-gression coefficients for the most important LiDAR-derived predictors according to therelative VIP indicated a negative relation between Pmass and the fractional cover of treeslarger than 6m (fcover_6_60_mean) (Figs 7 8) Moreover important LiDAR-derivedvariables indicated a negative relation of Pmass to the stand height (max_h_meanperc90_h_mean mean_h_mean) (Figs 7 8) Additionally fcover_2_6_mean relatedto the cover of shrubs was the most important variable in Pmass models using LiDAR-derived variables solely (Fig 8) Important hyperspectral bands were distributed acrossthe whole spectrum with a pronounced peak around 730 nm (Fig 4) The permutationof the calibration data according to the main forest types revealed that the success ofNmass and Pmass models was strongly dependent on two forest types being included (Fig9) Nmass models showed poor predictive performances when P sylvestris stands werenot included in the calibration dataset Similarly the absence of P serotina dominatedstands resulted in poor predictive performance of Pmass models This observation wasconsistent regardless of whether hyperspectral or LiDAR data were used as predictorsAdditionally model performances were strongly influenced by the variance in canopy

9

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

150 175 200 225 250

150

175

200

225

250

150

175

200

225

250

150

175

200

225

250

Predicted [gkg]

Obs

erve

d [g

kg]

Nmass

HS

HS

amp LiD

AR

LiDA

R

08 12 16 20

100

125

150

175

100

125

150

175

100

125

150

175

Predicted [gkg]

Pmass

Forest Type

C betulus

F sylvatica

Mixed broadleaf

P serotina

P sylvestris

Q robur

Figure 2 Mean predicted values resulting from 200 model predictions displayed against observedvalues for Nmass and Pmass of 50 field plots Error bars represent lower and upper quantiles of thepredictions The figures show results from models using hyperspectral bands (HS top) LiDAR-derivedpredictors (LiDAR bottom) and a combination of both (HS amp LiDAR middle) The coloringhighlights different forest types represented by dominant tree species

height and gap fraction of field plots included in each data permutation (Fig 10)Pmass models performed better with increasing variance in both structural propertiesThis contrasted with Nmass where the performance of imaging spectroscopy models de-creased with increasing variation in canopy height and gap fraction The performance ofNmass models was less affected by structural variation when including LiDAR-derivedvariables (Fig 10)

10

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

00

05

10

15

20

00

05

10

15

20

Wavelength [nm]

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 3 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Nmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD)

Mea

n V

IP (H

S)

Mea

n V

IP (H

S amp

LiD

AR

)

500 1000 1500 2000 2500

05

10

15

20

05

10

15

20

Wavelength [nm]

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 4 Mean VIP (Variable Importance in Projection) values of hyperspectral bands andLiDAR-derived variables resulting from 200 PLSR models for the prediction of Pmass The top panel isshowing the results from models using hyperspectral bands only bottom panels display results frommodels using a combination of hyperspectral bands and LiDAR-derived predictors Gray areas indicatethe range between the 10th and the 90th percentiles The bottom right panel is displaying mean VIPvalues of used LIDAR variables For simplification LIDAR-derived variables were grouped into fourclasses representing the vegetation cover (Fractional cover) the horizontal variability of vegetationcover (Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDARheight metrics (Height SD)

11

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus004

minus002

000

002

minus004

minus002

000

002

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

NmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 5 Mean PLSR Coefficients of hyperspectral bands and LiDAR-variables resulting from 200model calculations for predicting Nmass The top panel is showing the results from models usinghyperspectral bands only bottom panels display results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10th andthe 90th percentile The bottom right panel is displaying mean PLSR coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

HS

HS

amp L

iDA

R

500 1000 1500 2000 2500

minus010

minus005

000

005

010

minus010

minus005

000

005

010

Wavelength [nm]

Mea

n P

LSR

Coe

ffici

ent

PmassLiDAR variable group

Fractional cover

Fractional cover SD

Height

Height SD

LiDAR variables

Figure 6 Mean PLSR Coefficients of hyperspectral bands and LiDAR-derived variables resultingfrom 200 model calculations for predicting Pmass The top panel is showing the results from modelsusing hyperspectral bands only bottom panels display the results from models using a combination ofhyperspectral bands and LiDAR-derived variables Gray areas indicate the range between the 10 thand the 90 th percentile The bottom right panel is displaying mean PLSR Coefficients of usedLiDAR-derived variables For simplification LiDAR-derived variables were grouped into four classesrepresenting the vegetation cover (Fractional cover) the horizontal variability of vegetation cover(Fractional cover SD) LiDAR height metrics (Height) and the horizontal variability of LiDAR heightmetrics (Height SD) LiDAR variables are displayed in ascending order by variable importance

12

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

perc50_h_sd

mean_h_sd

perc90_h_sd

max_h_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_2_6_sd

fcover_6_60_sd

fcover_6_10_sd

00 05 10 15 20

Mean VIP

Nmass

minus005 minus004 minus003 minus002 minus001 000

Mean PLSR Coefficient

Nmass

fcover_05_60_mean

fcover_2_60_mean

perc10_h_mean

perc25_h_mean

perc75_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

max_h_mean

fcover_6_60_mean

00 05 10 15 20

Mean VIP

Pmass

minus0100 minus0075 minus0050 minus0025

Mean PLSR Coefficient

Pmass

Figure 7 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using a combination of hyperspectral bands and LiDAR-derived as predictorsError bars indicate the range between the 10th and 90th percentile

4 DiscussionIn this study we showed that LiDAR-derived information on canopy structure improvedpredictions of Nmass and Pmass based imaging spectroscopy instructurally heterogeneousforest stands This finding is in accordance with previous studies using optical remotesensing data which report a strong contribution of NIR reflectance for the prediction ofNmass in forest canopies (eg Martin et al 2008 Ollinger et al 2008 Wang et al 2016))that can be attributed to canopy structural properties (Knyazikhin et al 2013 Ollinger2011) Similarly Badgley et al (2017) found gross primary production on a global levelto be strongly related to structure-sensitive NIR reflectance These results point at theexistence of functional links between the biochemical and structural composition of forestcanopiesAn ecological explanation of such linkages follows from the economic theory (Bloom

et al 1985) The economic theory states that investments in the photosynthetic ma-

13

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

perc75_h_sd

cv_h_mean

fcover_05_2_sd

max_h_sd

fcover_6_10_sd

fcover_10_20_sd

fcover_2_60_sd

fcover_05_60_sd

fcover_6_60_sd

fcover_2_6_sd

00 05 10 15

Mean VIP

Nmass

minus0050 minus0025 0000 0025 0050

Mean PLSR Coefficient

Nmass

sd_h_mean

perc25_h_mean

perc10_h_mean

perc50_h_mean

mean_h_mean

perc90_h_mean

perc75_h_mean

max_h_mean

fcover_6_60_mean

fcover_2_6_mean

00 05 10 15

Mean VIP

Pmass

minus010 minus005 000 005

Mean PLSR Coefficient

Pmass

Figure 8 Mean VIP values (left) and mean PLSR coefficients (right) resulting from 200 PLSR modelsfor the prediction of Nmass (top) and Pmass (bottom) for the ten most important LiDAR-derivedvariables in models using LiDAR-derived predictors only Error bars indicate the range between the10th and 90th percentiles

chinery of plants will be realized only when the benefit of these investments exceedsthe anticipated costs As a result plant traits with small cost-to-benefit relationshipare favored under resource limitation leading to a functional convergence of structuraland physiological traits At the leaf level for example the negative correlation betweenleaf mass per area and leaf nitrogen or phosphorus concentration can be attributed tofunctional convergence (Diacuteaz et al 2016 Shipley et al 2006 Wright et al 2004) Eco-logical theory suggests that similar to the leaf level functional convergence can alsobe expected at the canopy level (Field 1991) leading to linkages between structuraland biochemical canopy properties In temperate and boreal forest ecosystems linksbetween structure and biochemistry are expressed at both the leaf and the canopy levelFor example broadleaved and coniferous trees show notable structural differences at thecanopy level which are expressed in different crown geometry branching architectureand leaf angle distribution (Ollinger 2011) Both leaf and canopy structural prop-

14

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

n = 47

n = 33

n = 41

n = 44

n = 46

n = 39

n = 47n =33 n = 41 n = 44

n = 46

n = 39Nmass Pmass

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

C betulus

F sylv

atica

Mixed broadleaf

P serotina

P sylve

stris

Q robur

00

02

04

06

08

Forest type excluded from model calibration

R2

cv HS

HS amp LiDAR

LiDAR

Figure 9 Predictive performances of Nmass and Pmass models using permuted calibration datasetsaccording to occurring forest types In each data permutation one forest type was excluded from thecalibration dataset Numbers above the bars represent the number of field plots included in eachcalibration dataset HS models using hyperspectral data HS amp LiDAR models using a combination ofhyperspectral and LiDAR data LiDAR models using LiDAR data only

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

08

00

02

04

06

Variance

R2

cv

Nmass

Gap fraction Height

HS

HS

amp LiD

AR

003 004 005 006 40 50 60 70

00

02

04

06

02

04

06

08

Variance

R2

cv

Pmass

Rsup2 = 001p = 012

Rsup2 = 011p lt 0001

Rsup2 = 019p lt 0001

Rsup2 = 004p = 0002

Rsup2 = 016p lt 0001

Rsup2 = 026p lt 0001

Rsup2 = 033p lt 0001

Rsup2 = 036p lt 0001

Figure 10 Predictive performances of Nmass and Pmass models depending on the variance of canopygap fraction and canopy height included in the calibration dataset Points represent the results from200 model repetitions using permuted calibration data Lines and values in each panel represent resultsfrom univariate linear regression between displayed variables Top panels are showing the results frommodels using imaging spectroscopy data (HS) only bottom panels the results from models using acombination of imaging spectroscopy and LiDAR data

erties have shown to influence spectral reflectance in similar ways resulting in higherreflectance of broadleaved canopies (Knyazikhin et al 2013 Ollinger 2011) At thesame time broadleaved trees are characterized by higher Nmass compared to conifer-ous tree species (Guumlsewell 2004 Han et al 2005 McNeil et al 2008 Serbin et al

15

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

Figure 11 Map sections showing forest types represented by their dominant tree species a canopyheight model (both in the middle) and median predictions of canopy level Nmass (top) and Pmass

(bottom) from models using hyperspectral bands (HS) LiDAR-derived predictors (LiDAR) or acombination of both (HS+LiDAR)

2014) Furthermore case studies show that forest canopy Nmass or Pmass can be alsorelated to other structural properties such as stand density above ground biomass orcrown-closure (Craven et al 2015 Goumlkkaya et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015)In the specific context of this study the success of Nmass and Pmass predictions was

strongly dependent on the presence of two forest types that exhibited biochemical andstructural differences compared to the co-occuring forest types Nmass predictions de-pended on the presence of P sylvestris stands in the calibration dataset Pinus sylvestriswas the only coniferous species in our study and was characterized by significantly lowerNmass than all other species In contrast Pmass predictions were mainly driven by P

16

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

serotina which was the species characterized by the highest Pmass concentrations inour study area Structural differences between P serotina and the other tree species inour study area mainly arise from its growth strategy and habitat preferences Prunusserotina is an early successional tree species with significantly smaller growth heightscompared to other tree species predominant in our study area Prunus serotina is often adominant species in young stands and often occurs in mature stands with sparse canopiesor in canopy gaps Our results suggest that species differences in structural andor opti-cal properties can serve as a surrogate to predict canopy chemistry using remote sensingat least across small study extents where differences in leaf nutrient concentrations canoften be explained by differences between species (Craven et al 2015 McNeil et al2008) For larger environmental gradients differences between species are often super-imposed by the high intra-specific variability of leaf biochemicals (Asner et al 2012Mellert and Goumlttlein 2012 Vilagrave-Cabrera et al 2015) which respond to strong vari-ation in climate and soil properties (Sardans et al 2015 Sun et al 2015) The factthat our results were strongly dependent on the occurrence of two species is limitingthe transferability of our findings to other study areas or broader spatial extents How-ever functional differences (eg between broadleaved and coniferous species or betweenearly and late successional species) that are manifested in structural and biochemicalproperties (Craven et al 2015 Kusumoto et al 2015 Sardans and Pentildeuelas 2015Vilagrave-Cabrera et al 2015) suggest that canopy structure can serve as a surrogate forpredicting biochemical properties also in different study contexts

Mapping Nmass

Predicting forest canopy Nmass using imaging spectroscopy has a long history Com-pared to previous studies which often report good (eg Smith et al 2003 Townsendet al 2003 Wang et al 2016) or even excellent (eg Martin et al 2008 Singh et al2015) predictive performances our models performed poorly We attribute this mainlyto the high structural diversity of the forest stands used for model calibration This highstructural diversity was for example expressed by strong variation of LAI values evenwithin stands of the same forest type (ie ranging from 18 to 61 for F sylvatica stands)Canopy structure strongly affects reflectance (Gerard and North 1997 Rautiainen et al2004) and a high variability in LAI has been found to hamper predictions of leaf biochem-istry at the canopy level (Asner and Martin 2008) Furthermore we included stands ofdifferent age classes with canopy heights ranging between 2 and 40 meters which alsoincreases variation in canopy reflectance especially in the VIS (Roberts et al 2004)Our results suggest that including LiDAR data can help to diminish these effects ofstructural heterogeneity when mapping Nmass (see Fig 10)In part the weak predictive performance of our Nmass models can be attributed to

the relatively low data range of Nmass in our study area (cf Asner et al 2015 Huberet al 2008 Martin et al 2008 Singh et al 2015 Smith et al 2003 Wang et al 2016)The range was especially low for all broadleaved species with no significant differencesbetween the two main species (F sylvatica Q robur) which were predominant in 36 of50 field plots (including mixed broadleaf) Furthermore the weak model performance

17

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

can presumably also be attributed to the usage of mass related nitrogen measuresbecause spectral reflectance is more closely linked to leaf biochemistry on an area basis(Grossman et al 1996 Roelofsen et al 2014)Furthermore the performance of the Nmass models may also be explained by the fact

that image acquisition and leaf sampling were from different years Although previousstudies suggest that there is only low variation of Nmass in temperate forest speciesbetween two consecutive years and during one growing season (McKown et al 2013Niinemets 2016 Reich et al 1991 Smith et al 2003) we cannot be 100 sure thatrelative differences between the species in our study area were stable between the yearsFajardo and Siefert (2016) found different patterns in Nmass between coniferous andbroad leaf species in the course of one growing season However they also found thatoverall species rankings concerning Nmass were stable throughout a growing seasonThe most important spectral bands selected in our Nmass models were situated in the

visible part of the spectrum A high contribution of the VIS region for Nmass predictionusing imaging spectroscopy was also observed by Asner et al (2015) and Singh et al(2015) In our study the importance of bands in the VIS can be attributed to differencesin reflectance between coniferous and broadleaved forest stands in this spectral region(see supplementary material Fig S4) These differences may arise from light absorptionof chlorophyll but may also be due to other leaf pigments like carotenoids and antho-cyanins that also have absorption characteristics in the VIS (Ollinger 2011 Ustin et al2009) Moreover structural canopy properties such as LAI or leaf angle distribution alsoinfluence reflectance in the VIS albeit to a lower extent than leaf pigments (Jacquemoudet al 2009) This is in accordance to previous studies that report the importance of theVIS region to discriminate between species (Fassnacht et al 2016 Roberts et al 2004)VIP values indicated only a minor contribution of spectral bands located in the NIR

and SWIR which is contrary to results of previous studies using image spectroscopy(Homolovaacute et al 2013) According to Ollinger (2011) NIR reflectance is especiallyimportant in datasets with only little variance in the VIS reflectance The high variancein the VIS reflectance (see supplementary material Fig S3) observed in this study maythus be an explanation for the minor contribution of NIR and SWIR bands Additionallyany signal in the infrared reflectance may be strongly disturbed by the high variabilityof canopy gaps in the field plots used for this study (Ollinger 2011)For mapping Nmass important LiDAR-derived variables were mainly connected to the

horizontal variation of canopy cover (fcover_6_10_sd fcover_6_60_sd fcover_2_6_sd)These three variables represent the variation of the fractional vegetation cover betweendifferent height thresholds in one 24m times 24m pixel They can thus be interpreted asindicators for spatial heterogeneity of the canopy The most important LiDAR-derivedvariable for predicting canopy level Nmass was the spatial variation of fractional veg-etation cover between 6 and 10m height (fcover_6_10_sd) which is related to theoccurrence of shrubs or small trees in the understory Low values either indicate littlevegetation present between 6 and 10m height as it can be observed in mature foreststands with closed canopies or very dense homogeneous vegetation as it can be observedin earlier successional stages High values indicate heterogeneous typically old-grownforest stands with gaps that are filled by young trees Similarly fcover_6_60_sd is

18

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

related to the horizontal heterogeneity of the tree canopy cover that was highest in Psylvestris stands (supplementary material Fig S5) Moreover fcover_2_6 also was high-est in P sylvestris stands indicating that LiDAR-derived variables helped to accentuatedifferences in Nmass between P sylvestris and broadleaved species

In summary Nmass predictions were strongly dependent on the presence of the onlyconiferous tree species P sylvestris Stands of P sylvestris were characterized by lowerNmass and higher spatial variation of canopy cover compared to broadleaved foreststands These structural differences could be well captured by LiDAR data (supple-mentary material Fig S5) Hence integrating LiDAR-derived information improvedmodels based on imaging spectroscopy data solely The poor performance of modelsusing hyperspectral data solely can be attributed to the high structural heterogeneityin the study area in terms of LAI and stand ages Our results suggest that LiDARdata can help to diminish the effect of canopy heterogeneity when mapping forest Nmass

using imaging spectroscopy

Mapping Pmass

Mapping leaf phosphorus with remote sensing has received much less attention comparedto Nmass Earlier mapping attempts were based on hyperspectral indices (Mirik et al2005) radiative transfer models (Porder et al 2005) and empirical models (Asner et al2015 Goumlkkaya et al 2015) Goumlkkaya et al (2015) achieved excellent predictive perfor-mances mapping Pmass in a boreal mixed forest using Hyperion imaging spectroscopydata Asner et al (2015) successfully mapped Pmass along a broad environmental gra-dient using airborne hyperspectral data and partial least squares regression Contraryto Nmass Pmass has no characteristic absorption features in the used wavelength rangeand thus the success of mapping Pmass can be rather attributed to correlations to othercanopy properties For many plant species Pmass is positively correlated with Nmass

(Elser et al 2010 Guumlsewell 2004) or leaf mass per area (Wright et al 2004) Fortemperate tree species Sardans et al (2015) found a negative correlation between aboveground biomass and leaf NP ratio due to higher P retention with increasing ageImportant bands for the prediction of Pmass were located throughout the whole range

of the spectra Asner et al (2015) and Goumlkkaya et al (2015) found similar results withimportant bands located in the VIS SWIR and NIR regions The most important se-lected LiDAR-derived variables were related to the cover of shrubs and the cover of trees(fcover_2_6_mean fcover_6_60_mean) While the shrub cover was positively relatedto Pmass tree canopy cover had an negative relationship both indicating higher Pmass invery young and very open stands We furthermore observed a negative relation betweenPmass and LiDAR-derived variables related to vegetation height (eg max_h_meanperc90_h_mean mean_h_mean) These variables are correlated to the mean height ofall LiDAR vegetation points and indicate that taller stands are related to lower PmassThe observation of higher Pmass in younger stands reflects the observation that earliersuccessional stages are often characterized by higher Pmass (Chai et al 2015 Eichenberget al 2015) Relations between important LiDAR-derived variables and Pmass can alsobe well explained by species-specific differences within the study area Prunus serotina

19

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

for which we observed highest Pmass values is a characteristic species of young and early-succesional stands in the forest of Compiegravegne The observed negative relation betweencanopy cover and Pmass can also be explained by species -specific differences particu-larly between P serotina P sylvestris and F sylvatica (see supplementary materialFig S5) Fagus sylvatica for which we observed smallest Pmass is forming most densecanopies in Mid-Europe while P sylvestris characterized by higher Pmass than most ofthe native broadleaved species is forming very sparse canopies Prunus serotina mostfrequently occurred in forest stands with sparse canopy cover and good light conditions(Starfinger et al 2003)In summary Pmass predictions were driven by one tree species occurring in young or

open forest stands Existing covariation between canopy structure and Pmass was bettercaptured by LiDAR data than by imaging spectroscopy The relative importance ofstructural properties for mapping Pmass is not surprising as phosphorus is not expectedto be directly represented in the spectral signal of plant canopies

5 ConclusionIn this study we used a combination of imaging spectroscopy and airborne LiDAR datafor mapping canopy Nmass and Pmass in a forest characterized by a high structuralheterogeneity For both Nmass and Pmass LiDAR-derived variables improved predic-tions based on imaging spectroscopy solely This highlights the importance of structuralproperties for remote sensing of biochemical variation in forest canopies For Nmass thepoor performance of hyperspectral data alone can be attributed to the high structuralheterogeneity in the study area in terms of LAI and stand ages LiDAR data helpedto capture this heterogeneity and hence improve model performances Both Nmass andPmass results were strongly influenced by the presence of only two tree species featuringstructural and biochemical properties different from their co-occurring tree species Thislimits the transferability of identified linkages between canopy structure and biochem-istry to other study settings However in the case of Nmass the known covariation withstructural properties existing at the leaf and canopy level suggests that canopy struc-ture used as proxy can support the mapping of Nmass also for different study settingsInformation on canopy structure derived from airborne LiDAR can help to understandexisting functional links

AcknowledgementsThis study is part of the project DIARS (Detection of invasive plant species and as-sessment of their impact on ecosystem properties through remote sensing) funded bythe ERA-Net BiodivERsA with the national funders ANR (Agence Nationale dela Recherche) BelSPO (Belgian Federal Science Policy Office) and DFG (DeutscheForschungsgemeinschaft) Michael Ewald is funded through the DFG research grantSCHM 21539-1 The authors would like to thank the Office National des Forecircts forgranting permission for leaf sampling and for providing the airborne LiDAR data We

20

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

also wish to thank Luc Croiseacute Fabien Spicher Anthony Viaud and Jens Warrie for theirhelp during field work Finally we would like to thank the anonymous reviewers for theirdetailed feedback which greatly helped to improve earlier versions of the manuscript

ReferencesAerts R Ewald M Nicolas M Piat J Skowronek S Lenoir J Hattab TGarzoacuten-Loacutepez CX Feilhauer H Schmidtlein S Rocchini D Decocq G SomersB Van De Kerchove R Denef K Honnay O 2017 Invasion by the alien treePrunus serotina alters ecosystem functions in a temperate deciduous forest Frontiersin Plant Science 8 179 doi103389fpls201700179

Ali AM Darvishzadeh R Skidmore AK van Duren I 2016 Effects of canopystructural variables on retrieval of leaf dry matter content and specific leaf area fromremotely sensed data IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing 9 898ndash909 doi101109JSTARS20152450762

Asner GP 1998 Biophysical and biochemical sources of variability in canopy re-flectance Remote Sensing of Environment 64 234ndash253 doi101016S0034-4257(98)00014-5

Asner GP Martin RE 2008 Spectral and chemical analysis of tropical forestsScaling from leaf to canopy levels Remote Sensing of Environment 112 3958ndash3970doi101016jrse200807003

Asner GP Martin RE Anderson CB Knapp DE 2015 Quantifying forestcanopy traits Imaging spectroscopy versus field survey Remote Sensing of Environ-ment 158 15ndash27 doi101016jrse201411011

Asner GP Martin RE Suhaili AB 2012 Sources of canopy chemical and spectraldiversity in lowland bornean forest Ecosystems 15 504ndash517 doi101007s10021-012-9526-2

Badgley G Field CB Berry JA 2017 Canopy near-infrared reflectance and terres-trial photosynthesis Science Advances 3 e1602244 doi101126sciadv1602244

Bloom AJ Chapin III F Mooney HA 1985 resource limitation in plants-aneconomic analogy Annual Review of Ecology and Systematics 16 363ndash392 doi101146annureves16110185002051

Cade BS 1997 Comparison of tree basal area and canopy cover in habitat modelssubalpine forest The Journal of Wildlife Management 61 326ndash335 doi1023073802588

Chabrerie O Verheyen K Saguez R Decocq G 2008 Disentangling relationshipsbetween habitat conditions disturbance history plant diversity and American black

21

cherry Prunus serotina Ehrh) invasion in a European temperate forest Diversity andDistributions 14 204ndash212 doi101111j1472-4642200700453x

Chai Y Yue M Wang M Xu J Liu X Zhang R Wan P 2015 Plant functionaltraits suggest a change in novel ecological strategies for dominant species in the stagesof forest succession Oecologia 180 771ndash783 doi101007s00442-015-3483-3

Chapin FS 2003 Effects of plant traits on ecosystem and regional processes a con-ceptual framework for predicting the consequences of global change Annals of Botany91 455ndash463 doi101093aobmcg041

Closset-Kopp D Saguez R Decocq G 2010 Differential growth patterns and fitnessmay explain contrasted performances of the invasive Prunus serotina in its exoticrange Biological Invasions 13 1341ndash1355 doi101007s10530-010-9893-6

Craven D Hall JS Berlyn GP Ashton MS Breugel Mv 2015 Changing gearsduring succession shifting functional strategies in young tropical secondary forestsOecologia 179 293ndash305 doi101007s00442-015-3339-x

Curran PJ Kupiec JA Smith GM 1997 Remote sensing the biochemical compo-sition of a slash pine canopy IEEE Transactions on Geoscience and Remote Sensing35 415ndash420 doi10110936563280

Dahlin KM Asner GP Field CB 2013 Environmental and community controlson plant canopy chemistry in a Mediterranean-type ecosystem Proceedings of theNational Academy of Sciences 110 6895ndash6900 doi101073pnas1215513110

Di Palo F Fornara D 2015 Soil fertility and the carbonnutrient stoichiometry ofherbaceous plant species Ecosphere 6 1ndash15 doi101890ES15-004511

Diacuteaz S Kattge J Cornelissen JHC Wright IJ Lavorel S Dray S Reu BKleyer M Wirth C Colin Prentice I Garnier E Boumlnisch G Westoby MPoorter H Reich PB Moles AT Dickie J Gillison AN Zanne AE ChaveJ Joseph Wright S Sheremetrsquoev SN Jactel H Baraloto C Cerabolini BPierce S Shipley B Kirkup D Casanoves F Joswig JS Guumlnther A FalczukV Ruumlger N Mahecha MD Gorneacute LD 2016 The global spectrum of plant formand function Nature 529 167ndash171 doi101038nature16489

Eichenberg D Trogisch S Huang Y He JS Bruelheide H 2015 Shifts in com-munity leaf functional traits are related to litter decomposition along a secondaryforest succession series in subtropical China Journal of Plant Ecology 8 401ndash410doi101093jpertu021

Elser JJ Fagan WF Kerkhoff AJ Swenson NG Enquist BJ 2010 Bi-ological stoichiometry of plant production metabolism scaling and ecological re-sponse to global change New Phytologist 186 593ndash608 doi101111j1469-8137201003214x

22

Ewald M Dupke C Heurich M Muumlller J Reineking B 2014 LiDAR remotesensing of forest structure and GPS telemetry data provide insights on winter habitatselection of european roe deer Forests 5 1374ndash1390 doi103390f5061374

Fajardo A Siefert A 2016 Phenological variation of leaf functional traits withinspecies Oecologia 180 951ndash959 doi101007s00442-016-3545-1

Fassnacht FE Hartig F Latifi H Berger C Hernaacutendez J Corvalaacuten P Koch B2014 Importance of sample size data type and prediction method for remote sensing-based estimations of aboveground forest biomass Remote Sensing of Environment154 102ndash114 doi101016jrse201407028

Fassnacht FE Latifi H Stereńczak K Modzelewska A Lefsky M Waser LTStraub C Ghosh A 2016 Review of studies on tree species classification fromremotely sensed data Remote Sensing of Environment 186 64ndash87 doi101016jrse201608013

Feilhauer H Asner GP Martin RE Schmidtlein S 2010 Brightness-normalizedpartial least squares regression for hyperspectral data Journal of Quantitative Spec-troscopy and Radiative Transfer 111 1947ndash1957 doi101016jjqsrt201003007

Garnier E Lavorel S Ansquer P Castro H Cruz P Dolezal J Eriksson OFortunel C Freitas H Golodets C Grigulis K Jouany C Kazakou E KigelJ Kleyer M Lehsten V Lepš J Meier T Pakeman R Papadimitriou MPapanastasis VP Quested H Queacutetier F Robson M Roumet C Rusch GSkarpe C Sternberg M Theau JP Theacutebault A Vile D Zarovali MP 2007Assessing the effects of land-use change on plant traits communities and ecosystemfunctioning in grasslands a standardized methodology and lessons from an applicationto 11 European sites Annals of Botany 99 967ndash985 doi101093aobmcl215

Gerard FF North PRJ 1997 Analyzing the effect of structural variability andcanopy gaps on forest BRDF using a geometric-optical model Remote Sensing ofEnvironment 62 46ndash62 doi101016S0034-4257(97)00070-9

Gill SJ Biging GS Murphy EC 2000 Modeling conifer tree crown radius andestimating canopy cover Forest Ecology and Management 126 405ndash416 doi101016S0378-1127(99)00113-9

Grassi G Vicinelli E Ponti F Cantoni L Magnani F 2005 Seasonal and inter-annual variability of photosynthetic capacity in relation to leaf nitrogen in a deciduousforest plantation in northern Italy Tree Physiology 25 349ndash360

Grossman YL Ustin SL Jacquemoud S Sanderson EW Schmuck G Verde-bout J 1996 Critique of stepwise multiple linear regression for the extraction of leafbiochemistry information from leaf reflectance data Remote Sensing of Environment56 182ndash193 doi1010160034-4257(95)00235-9

23

Goumlkkaya K Thomas V Noland TL McCaughey H Morrison I Treitz P2015 Prediction of macronutrients at the canopy level using spaceborne imagingspectroscopy and LiDAR data in a mixedwood boreal forest Remote Sensing 79045ndash9069 doi103390rs70709045

Guumlsewell S 2004 N P ratios in terrestrial plants variation and functional significanceNew Phytologist 164 243ndash266 doi101111j1469-8137200401192x

Han W Fang J Guo D Zhang Y 2005 Leaf nitrogen and phosphorus stoichiom-etry across 753 terrestrial plant species in China New Phytologist 168 377ndash385doi101111j1469-8137200501530x

Homolovaacute L Malenovskyacute Z Clevers JGPW Garciacutea-Santos G Schaepman ME2013 Review of optical-based remote sensing for plant trait mapping EcologicalComplexity 15 1ndash16 doi101016jecocom201306003

Huber S Kneubuumlhler M Psomas A Itten K Zimmermann NE 2008 Estimatingfoliar biochemistry from hyperspectral data in mixed forest canopy Forest Ecologyand Management 256 491ndash501 doi101016jforeco200805011

Jacquemoud S Verhoef W Baret F Bacour C Zarco-Tejada PJ Asner GPFranccedilois C Ustin SL 2009 PROSPECT+SAIL models A review of use forvegetation characterization Remote Sensing of Environment 113 S56ndashS66 doi101016jrse200801026

Jonard M Fuumlrst A Verstraeten A Thimonier A Timmermann V Potočić NWaldner P Benham S Hansen K Merilauml P Ponette Q de la Cruz ACRoskams P Nicolas M Croiseacute L Ingerslev M Matteucci G Decinti B Basci-etto M Rautio P 2015 Tree mineral nutrition is deteriorating in Europe GlobalChange Biology 21 418ndash430 doi101111gcb12657

Kimberley A Alan Blackburn G Duncan Whyatt J Smart SM 2014 Traitsof plant communities in fragmented forests the relative influence of habitat spatialconfiguration and local abiotic conditions Journal of Ecology 102 632ndash640 doi1011111365-274512222

Knyazikhin Y Schull MA Stenberg P Motildettus M Rautiainen M Yang Y Mar-shak A Carmona PL Kaufmann RK Lewis P Disney MI Vanderbilt VDavis AB Baret F Jacquemoud S Lyapustin A Myneni RB 2013 Hy-perspectral remote sensing of foliar nitrogen content Proceedings of the NationalAcademy of Sciences 110 E185ndashE192 doi101073pnas1210196109

Kokaly RF Asner GP Ollinger SV Martin ME Wessman CA 2009 Charac-terizing canopy biochemistry from imaging spectroscopy and its application to ecosys-tem studies Remote Sensing of Environment 113 78ndash91 doi101016jrse200810018

24

Korhonen L Korpela I Heiskanen J Maltamo M 2011 Airborne discrete-returnLIDAR data in the estimation of vertical canopy cover angular canopy closure andleaf area index Remote Sensing of Environment 115 1065ndash1080 doi101016jrse201012011

Kusumoto B Shiono T Miyoshi M Maeshiro R Fujii Sj Kuuluvainen T Kub-ota Y 2015 Functional response of plant communities to clearcutting managementimpacts differ between forest vegetation zones Journal of Applied Ecology 52 171ndash180 doi1011111365-266412367

Lamarque P Lavorel S Mouchet M Queacutetier F 2014 Plant trait-based modelsidentify direct and indirect effects of climate change on bundles of grassland ecosystemservices Proceedings of the National Academy of Sciences 111 13751ndash13756 doi101073pnas1216051111

Lavorel S Grigulis K Lamarque P Colace MP Garden D Girel J PelletG Douzet R 2011 Using plant functional traits to understand the landscapedistribution of multiple ecosystem services Journal of Ecology 99 135ndash147 doi101111j1365-2745201001753x

Martin ME Aber JD 1997 High spectral resolution remote sensing of forest canopylignin nitrogen and ecosystem processes Ecological Applications 7 431ndash443 doi1023072269510

Martin ME Plourde LC Ollinger SV Smith ML McNeil BE 2008 A gen-eralizable method for remote sensing of canopy nitrogen across a wide range of forestecosystems Remote Sensing of Environment 112 3511ndash3519 doi101016jrse200804008

Masek JG Hayes DJ Joseph Hughes M Healey SP Turner DP 2015 Therole of remote sensing in process-scaling studies of managed forest ecosystems ForestEcology and Management 355 109ndash123 doi101016jforeco201505032

McKown AD Guy RD Azam MS Drewes EC Quamme LK 2013 Seasonalityand phenology alter functional leaf traits Oecologia 172 653ndash665 doi101007s00442-012-2531-5

McNeil BE Read JM Sullivan TJ McDonnell TC Fernandez IJ DriscollCT 2008 The spatial pattern of nitrogen cycling in the adirondack park New YorkEcological Applications 18 438ndash452 doi10189007-02761

Melillo JM Aber JD Muratore JF 1982 Nitrogen and lignin control of hardwoodleaf litter decomposition dynamics Ecology 63 621ndash626 doi1023071936780

Mellert KH Goumlttlein A 2012 Comparison of new foliar nutrient thresholds derivedfrom van den Burgrsquos literature compilation with established central European refer-ences European Journal of Forest Research 131 1461ndash1472 doi101007s10342-012-0615-8

25

Mirik M Norland JE Crabtree RL Biondini ME 2005 Hyperspectral one-meter-resolution remote sensing in Yellowstone National Park Wyoming I foragenutritional values Rangeland Ecology amp Management 58 452ndash458 doi10211104-171

Niinemets U 2016 Leaf age dependent changes in within-canopy variation in leaffunctional traits a meta-analysis Journal of Plant Research 129 313ndash338 doi101007s10265-016-0815-2

Ollinger SV 2011 Sources of variability in canopy reflectance and the convergentproperties of plants New Phytologist 189 375ndash394 doi101111j1469-8137201003536x

Ollinger SV Richardson AD Martin ME Hollinger DY Frolking SE ReichPB Plourde LC Katul GG Munger JW Oren R Smith ML U KTPBolstad PV Cook BD Day MC Martin TA Monson RK Schmid HP2008 Canopy nitrogen carbon assimilation and albedo in temperate and borealforests Functional relations and potential climate feedbacks Proceedings of the Na-tional Academy of Sciences 105 19336ndash19341 doi101073pnas0810021105

Ollinger SV Smith ML Martin ME Hallett RA Goodale CL Aber JD2002 Regional variation in foliar chemistry and N cycling among forests of di-verse history and composition Ecology 83 339ndash355 doi1018900012-9658(2002)083[0339RVIFCA]20CO2

Porder S Asner GP Vitousek PM 2005 Ground-based and remotely sensed nu-trient availability across a tropical landscape Proceedings of the National Academyof Sciences of the United States of America 102 10909ndash10912 doi101073pnas0504929102

Pullanagari RR Kereszturi G Yule IJ 2016 Mapping of macro and micro nutrientsof mixed pastures using airborne AisaFENIX hyperspectral imagery ISPRS Journalof Photogrammetry and Remote Sensing 117 1ndash10 doi101016jisprsjprs201603010

R Core Team 2016 R A Language and Environment for Statistical ComputingR Foundation for Statistical Computing Vienna Austria URL httpswwwR-projectorg

Rautiainen M Stenberg P Nilson T Kuusk A 2004 The effect of crown shapeon the reflectance of coniferous stands Remote Sensing of Environment 89 41ndash52doi101016jrse200310001

Reich PB 2012 Key canopy traits drive forest productivity Proceedings of the RoyalSociety of London B Biological Sciences 279 2128ndash2134 doi101098rspb20112270

26

Reich PB Walters MB Ellsworth DS 1991 Leaf age and season influence therelationships between leaf nitrogen leaf mass per area and photosynthesis in mapleand oak trees Plant Cell amp Environment 14 251ndash259 doi101111j1365-30401991tb01499x

Roberts DA Ustin SL Ogunjemiyo S Greenberg J Dobrowski SZ Chen JHinckley TM 2004 Spectral and structural measures of northwest forest vegetationat leaf to landscape scales Ecosystems 7 545ndash562 doi101007s10021-004-0144-5

Roelofsen HD van Bodegom PM Kooistra L Witte JPM 2014 Predicting leaftraits of herbaceous species from their spectral characteristics Ecology and Evolution4 706ndash719 doi101002ece3932

Sardans J Janssens IA Alonso R Veresoglou SD Rillig MC Sanders TGCarnicer J Filella I Farreacute-Armengol G Pentildeuelas J 2015 Foliar elementalcomposition of European forest tree species associated with evolutionary traits andpresent environmental and competitive conditions Global Ecology and Biogeography24 240ndash255 doi101111geb12253

Sardans J Pentildeuelas J 2015 Trees increase their PN ratio with size Global Ecologyand Biogeography 24 147ndash156 doi101111geb12231

Schlerf M Atzberger C Hill J Buddenbaum H Werner W Schuumller G 2010Retrieval of chlorophyll and nitrogen in Norway spruce (Picea abies L Karst) us-ing imaging spectroscopy International Journal of Applied Earth Observation andGeoinformation 12 17ndash26 doi101016jjag200908006

Schmidtlein S Feilhauer H Bruelheide H 2012 Mapping plant strategy types usingremote sensing Journal of Vegetation Science 23 395ndash405 doi101111j1654-1103201101370x

Serbin SP Singh A McNeil BE Kingdon CC Townsend PA 2014 Spectro-scopic determination of leaf morphological and biochemical traits for northern temper-ate and boreal tree species Ecological Applications 24 1651ndash1669 doi10189013-21101

Serrano L Pentildeuelas J Ustin SL 2002 Remote sensing of nitrogen and lignin inMediterranean vegetation from AVIRIS data Decomposing biochemical from struc-tural signals Remote Sensing of Environment 81 355ndash364 doi101016S0034-4257(02)00011-1

Shipley B Lechowicz MJ Wright I Reich PB 2006 Fundamental trade-offsgenerating the worldwide leaf economics spectrum Ecology 87 535ndash541 doi10189005-1051

Simic A Chen JM Noland TL 2011 Retrieval of forest chlorophyll content usingcanopy structure parameters derived from multi-angle data the measurement con-

27

cept of combining nadir hyperspectral and off-nadir multispectral data InternationalJournal of Remote Sensing 32 5621ndash5644 doi101080014311612010507257

Singh A Serbin SP McNeil BE Kingdon CC Townsend PA 2015 Imagingspectroscopy algorithms for mapping canopy foliar chemical and morphological traitsand their uncertainties Ecological Applications 25 2180ndash2197 doi10189014-20981

Smith ML Martin ME Plourde L Ollinger SV 2003 Analysis of hyperspectraldata for estimation of temperate forest canopy nitrogen concentration comparison be-tween an airborne (AVIRIS) and a spaceborne (Hyperion) sensor IEEE Transactionsof Geoscience and Remote Sensing 41 1332ndash1337

Starfinger U Kowarik I Rode M Schepker H 2003 From desirable ornamentalplant to pest to accepted addition to the flora ndash the perception of an alien treespecies through the centuries Biological Invasions 5 323ndash335 doi101023BBINV00000055731480007

Sterckx S Vreys K Biesemans J Iordache MD Bertels L Meuleman K 2016Atmospheric correction of APEX hyperspectral data Miscellanea Geographica 2016ndash20 doi101515mgrsd-2015-0022

Suding KN Lavorel S Chapin FS Cornelissen JHC Diacuteaz S Garnier EGoldberg D Hooper DU Jackson ST Navas ML 2008 Scaling environ-mental change through the community-level a trait-based response-and-effect frame-work for plants Global Change Biology 14 1125ndash1140 doi101111j1365-2486200801557x

Sun X Kang H Kattge J Gao Y Liu C 2015 Biogeographic patterns of multi-element stoichiometry of Quercus variabilis leaves across China Canadian Journal ofForest Research 45 1827ndash1834 doi101139cjfr-2015-0110

Talkner U Meiwes KJ Potočić N Seletković I Cools N Vos BD Rautio P2015 Phosphorus nutrition of beech (Fagus sylvatica L) is decreasing in EuropeAnnals of Forest Science 72 919ndash928 doi101007s13595-015-0459-8

Townsend PA Foster JR Chastain RA Currie WS 2003 Application of imagingspectroscopy to mapping canopy nitrogen in the forests of the central AppalachianMountains using Hyperion and AVIRIS IEEE Transactions on Geoscience and RemoteSensing 41 1347ndash1354 doi101109TGRS2003813205

Ustin SL Gitelson AA Jacquemoud S Schaepman M Asner GP Gamon JAZarco-Tejada P 2009 Retrieval of foliar information about plant pigment systemsfrom high resolution spectroscopy Remote Sensing of Environment 113 S67ndashS77doi101016jrse200810019

28

Vilagrave-Cabrera A Martiacutenez-Vilalta J Retana J 2015 Functional trait variation alongenvironmental gradients in temperate and Mediterranean trees Global Ecology andBiogeography 24 1377ndash1389 doi101111geb12379

Vreys K Iordache MD Biesemans J Meuleman K 2016 Geometric correction ofAPEX hyperspectral data Miscellanea Geographica 20 11ndash15 doi101515mgrsd-2016-0006

Wang Z Wang T Darvishzadeh R Skidmore AK Jones S Suarez L WoodgateW Heiden U Heurich M Hearne J 2016 Vegetation indices for mapping canopyfoliar nitrogen in a mixed temperate forest Remote Sensing 8 491 doi103390rs8060491

Wilson KB Baldocchi DD Hanson PJ 2000 Spatial and seasonal variability ofphotosynthetic parameters and their relationship to leaf nitrogen in a deciduous forestTree Physiology 20 565ndash578 doi101093treephys209565

Wright IJ Reich PB Westoby M Ackerly DD Baruch Z Bongers F Cavender-Bares J Chapin T Cornelissen JHC Diemer M Flexas J Garnier E GroomPK Gulias J Hikosaka K Lamont BB Lee T Lee W Lusk C MidgleyJJ Navas ML Niinemets U Oleksyn J Osada N Poorter H Poot P PriorL Pyankov VI Roumet C Thomas SC Tjoelker MG Veneklaas EJ VillarR 2004 The worldwide leaf economics spectrum Nature 428 821ndash827 doi101038nature02403

Xiao Y Zhao W Zhou D Gong H 2014 Sensitivity analysis of vegeta-tion reflectance to biochemical and biophysical variables at leaf canopy and re-gional scales IEEE Transactions on Geoscience and Remote Sensing 52 4014ndash4024doi101109TGRS20132278838

Zolkos SG Goetz SJ Dubayah R 2013 A meta-analysis of terrestrial abovegroundbiomass estimation using lidar remote sensing Remote Sensing of Environment 128289ndash298 doi101016jrse201210017

29