Nc

Oa

b

a

ARR2A

KCCNNR

I

lteettBpd

f

If

h0

International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

jo ur nal home page: www.elsev ier .com/ locate / jag

on-destructive estimation of foliar chlorophyll and carotenoidontents: Focus on informative spectral bands

z Kira a, Raphael Linker a, Anatoly Gitelson a,b,∗

Faculty of Civil and Environmental Engineering, Israel Institute of Technology, Technion, Technion City, Haifa, IsraelSchool of Natural Resources, University of Nebraska, Lincoln, USA

r t i c l e i n f o

rticle history:eceived 30 September 2014eceived in revised form8 December 2014ccepted 7 January 2015

eywords:arotenoidshlorophylleural networkon-destructive techniqueeflectance

a b s t r a c t

Leaf pigment content provides valuable insight into the productivity, physiological and phenologicalstatus of vegetation. Measurement of spectral reflectance offers a fast, nondestructive method for pigmentestimation. A number of methods were used previously for estimation of leaf pigment content, however,spectral bands employed varied widely among the models and data used. Our objective was to findinformative spectral bands in three types of models, vegetation indices (VI), neural network (NN) andpartial least squares (PLS) regression, for estimating leaf chlorophyll (Chl) and carotenoids (Car) contentsof three unrelated tree species and to assess the accuracy of the models using a minimal number of bands.The bands selected by PLS, NN and VIs were in close agreement and did not depend on the data used.The results of the uninformative variable elimination PLS approach, where the reliability parameter wasused as an indicator of the information contained in the spectral bands, confirmed the bands selected bythe VIs, NN, and PLS models. All three types of models were able to accurately estimate Chl content withcoefficient of variation below 12% for all three species with VI showing the best performance. NN and PLS

using reflectance in four spectral bands were able to estimate accurately Car content with coefficient ofvariation below 14%. The quantitative framework presented here offers a new way of estimating foliarpigment content not requiring model re-parameterization for different species. The approach was testedusing the spectral bands of the future Sentinel-2 satellite and the results of these simulations showedthat accurate pigment estimation from satellite would be possible.

© 2015 Elsevier B.V. All rights reserved.

ntroduction

Pigments are integrally related to the physiological function ofeaves. Chlorophylls (Chl) absorb light energy and transfer it intohe photosynthetic apparatus. Carotenoids (Car) can also contributenergy to the photosynthetic system. When incident light energyxceeds that needed for photosynthesis, the Car that composehe xanthophyll cycle dissipate excess energy avoiding damageo the photosynthetic system (Demmig-Adams and Adams, 1996).ecause of the importance of pigments for vegetation function,igment content may provide information concerning plant pro-

uctivity and physiological state.

There are several reasons why leaf pigmentation is importantrom an applied perspective to both land managers and ecophysiol-

∗ Corresponding author at: Faculty of Civil and Environmental Engineering, Israelnstitute of Technology, Technion, Technion City, Haifa, Israel. Tel.: +972 52 9544862;ax: +972 4 8295708.

E-mail address: agitelson2@unl.edu (A. Gitelson).

ttp://dx.doi.org/10.1016/j.jag.2015.01.003303-2434/© 2015 Elsevier B.V. All rights reserved.

ogists (Richardson et al., 2002). First, the amount of solar radiationabsorbed by a leaf is largely a function of the foliar contents ofphotosynthetic pigments, and therefore Chl content directly relatesto photosynthetic potential and hence primary production (Curranet al., 1990; Filella et al., 1995; Kergoat et al., 2008). Second, much ofleaf nitrogen is incorporated in Chl so quantifying Chl content givesan indirect measure of nutrient status (Filella et al., 1995; Moranet al., 2000; Baret et al., 2007; Kergoat et al., 2008; Schlemmeret al., 2013). Third, pigmentation can be directly related to stressphysiology, as contents of Car increase and contents of Chl gen-erally decrease under stress and during senescence (Penuelas andFilella, 1998). Fourth, the relative contents of pigments are knownto change with abiotic factors such as light (e.g., sun leaves havea higher Chl-a/Chl-b ratio; Larcher, 1995) and so quantifying theseproportions can provide important information about relationshipsbetween plants and their environment.

Traditional methods of pigment analysis, extraction and spec-trophotometric or HPLC measurement, require destruction of themeasured leaves and do not permit measurement of changes inpigments over time for a single leaf. These techniques are time

2 th Obs

caAosoamtNapqri(ae

pcepb(2rigtiavrdtaTa

cafcsmlc

coreitgee

pye(nrr

and maple) were used (Table 1). Pigment content varied widely:total Chl from 4 to 675 mg m−2 and total carotenoids from 16 to137.2 mg m−2.

Table 1Pigment content (in mg m−2) of maple, chestnut and beech leaves examined in thisstudy.

Chl-a Chl-b Total Chl Total Car

Maplen = 38

Median 181.7 62.0 242.5 45.9Average 192.6 69.5 262.1 48.4Minimum 2.8 1.0 4.0 16.0Maximum 419.9 150.1 570.0 82.6

Chestnut Median 51.3 15.5 66.7 49.6n = 18 Average 105.0 34.7 139.7 48.9

Minimum 7.0 3.0 10.0 26.0Maximum 335.9 134.5 470.4 83.2

52 O. Kira et al. / International Journal of Applied Ear

onsuming and expensive, thus making assessment of the over-ll vegetation state of landscapes and ecosystems impractical.lternative solutions of leaf pigment analysis are non-destructiveptical methods. Monitoring plant physiological status via mea-uring leaf reflectance possesses a number of distinct advantagesver traditional destructive approaches. The most important onesre simplicity, sensitivity, reliability and a high throughput. Theseethods are non-destructive, inexpensive, and quick and can

herefore be applied across spatial scales (Gamon and Qiu, 1999).on-destructive techniques save a great deal of manual labornd therefore have a great potential for application in studies oflant productivity, physiology and so on. Developing methods touantify pigment content and composition using non-destructiveemote reflectance measurement would clearly provide a capabil-ty that could advance understanding of photosynthetic processese.g., light regulation, photooxidation, chlorophyll fluorescence)nd insight into detection and monitoring of foliar condition (e.g.,nvironmental stressors).

A large number of reflectance-based methods have been pro-osed to detect plant pigments, ranging from empirical pigmentontent vs. reflectance relationships to radiative transfer mod-ls. Analytical radiative transfer models have the potential toroduce accurate and consistent prediction of pigment contentsecause they use the full spectrum rather than individual bandsJacquemoud and Baret, 1990; Maier et al., 1999; Feret et al.,008, 2011). Analytical models were successfully used to optimizeemotely sensed vegetation indices (VI) for estimating leaf chem-cal constituents (Feret et al., 2011). Synthetic reflectance spectraenerated by a radiative transfer model, PROSPECT-5, were usedo develop statistical relationships between leaf optical and chem-cal properties, which were applied to experimental data withoutny readjustment. Two methods used in remote sensing to estimateegetation chemical composition, VI and Partial Least Squares (PLS)egression, were trained both on the synthetic and experimentalatasets, and validated against observations. The study used syn-hetic data to establish several relationships to estimate leaf Chlnd Car content and validated these on a large variety of leaf types.he straightforward method described brought the possibility topply or adapt statistical relationships to any type of leaf.

The goodness-of-fit of radiative transfer models to predict opti-al properties of leaves or needles depends on how well understoodll processes affecting reflectance are, and how they are accountedor in the models (le Maire et al., 2004). However, these analyti-al models are difficult to invert and require information about leaftructure that may not be available or, if not accurate, result in poorodel performance. Consequently, most relationships between

eaf reflectance and pigment contents have been derived empiri-ally.

Among techniques using reflectance to quantify leaf pigmentontent are vegetation indices (VI) employing a few spectral bandsr multiple bands. These indices are based on knowledge of theeflectance properties of leaf biochemical components. le Mairet al. (2004) provided a comprehensive listing of the Chl spectralndices published until 2002. More complete reviews of the prac-ical and theoretical considerations of reflectance spectroscopy areiven by Curran et al. (1990), Gamon and Surfus, (1999), le Mairet al. (2004), Richardson et al. (2002), Blackburn, (2007a), Hatfieldt al. (2008), and Ustin et al. (2009).

Various processing approaches have been investigated, such asrincipal components analysis (Yao and Tian, 2003), factor anal-sis (Coops et al., 2002), stepwise multiple regression (O’Neillt al., 2002; Osborne et al., 2002), artificial neural networks (NN)

Tumbo et al., 2002; Chen et al., 2007), or wavelet-based tech-iques (Blackburn, 2007b), among others. Despite 30 years of leafeflectance spectroscopy research, it remains an area of activeesearch. In most cases the techniques for pigment estimation have

ervation and Geoinformation 38 (2015) 251–260

been tested for a single species or at most a few related species andthus it is not clear whether they can be applied across species withvarying leaf structural characteristics. In other cases one technique(e.g., VI) was applied for different species (e.g., Sims and Gamon,2002) with no comparison with other techniques. A key problemis the selection of optimal spectral bands and an appropriate pro-cessing approach for pigment estimation among the vast array ofthose available.

Our objective in this study was to test the performance of vege-tation indices, neural network (NN) and partial least squares (PLS)regression for estimating foliar Chl and Car contents of three unre-lated tree species. The attempt was made to find techniques that areinsensitive to species and leaf structure variation and thus could beapplied in large scale remote sensing studies without extensive cal-ibration. In particular, we focused on finding the most informativespectral bands, i.e., developing models that allow accurate estima-tion of pigment content with a minimum number of spectral bands.We found that all three techniques are very accurate for estimat-ing Chl content. Rededge chlorophyll index used two quite widespectral bands in rededge and NIR regions, while both NN and PLSused three 20 nm wide bands. NN and PLS were more accurate thancarotenoid reflectance index estimating Car content in all threespecies taken together, thus were not species-specific. Four 20 nmwide spectral bands centered at 510, 550, 720 and 770 nm allowedaccurate Car estimation in three species by single algorithm. Themost informative spectral bands found using uninformative vari-able elimination PLS (UVE PLS) coincided with spectral bands usedin three tested techniques, VI, NN and PLS. All three techniqueswere tested with reflectance simulated in spectral bands of nearfuture Sentinel-2 satellite and were found to be estimating accu-rately pigment content.

Methods

Pigment content and reflectance measurement

For estimating the Chl and Car contents, anthocyanin free juve-nile, mature and senescent leaves collected from 1992 to 2005were used. Norway maple and horse chestnut leaves were from apark at Moscow State University (Russia), beech leaves were fromthe University of Karlsruhe campus (Germany). The leaf total Chland Car content was determined analytically from the same leafsamples used for reflectance measurement (details are in Gitelsonet al., 2001, 2002, 2003). Pigment content was expressed on a leafarea basis. Three data sets containing 90 leaves (beech, chestnut,

Beech Median 256.9 95.7 352.6 90.2n = 34 Average 271.6 101.2 371.0 87.8

Minimum 63.5 21.9 85.5 30.1Maximum 541.4 192.0 675.0 137.2

th Obs

wst2r

D

andNwwofLpp

sPst2obbp

rCTba

to(otr

wPmt

c

wsamm

aetc2

C

O. Kira et al. / International Journal of Applied Ear

Adaxial reflectance (�) spectra of maple and chestnut leavesere taken in the spectral range between 400 and 800 nm with

pectral resolution about 1.5 nm with a Hitachi 150–20 spectropho-ometer and spectra of beech leaves were taken with a Shimadzu101 PC spectrophotometer in the range 400–800 nm with spectralesolution of 2 nm (details are in Gitelson et al., 2003).

ata analysis

Two common chemometric methods were used for the datanalysis: partial least squares (PLS) regression (Brereton, 2005) andeural networks (NN) (Haykin, 2005). Preliminary analysis (notetailed) led to selecting four factors for the PLS regression. TheN used in this study was a basic two-layer feed-forward networkith two hidden nodes. Both linear and sigmoid transfer functionsere tested and since similar results were obtained with both types

f transfer function, only the results obtained with the linear trans-er function are presented. The NN training was performed with theevenberg–Marquardt optimization algorithm, and stopped by theerformance error), minimal gradient, and Marquardt adjustmentarameter (1·1010).

Hyperspectral reflectance spectra were resampled with 10 nmpectral resolution in the range from 400 to 800 nm. NN andLS models were developed using non-overlapping 20 nm widepectral bands containing three contiguous measurements. Addi-ionally, models were developed on reflectance averaged over0 nm wide spectral bands. No significant differences werebserved in the results using reflectance averaged over 20 nm wideands and non-averaged and, therefore, only the results of analysesased on non-averaged reflectance (20 nm wide spectral bands) areresented.

In order to determine the number and position of bandsequired, an exhaustive search was conducted separately for bothhl and Car. All possible combinations of k bands were investigated.he number of bands k was 4 for Chl and 5 for Car. In both cases theest model included 3 or 4 bands, respectively, corresponding to 9nd 12 measurements.

To investigate the effect of the position of each spectral band onhe accuracy of pigment content estimation by NN and PLS models,ne spectral band was fixed and the positions of the other bandstwo for Chl and three for Car) were varied. Spectra of average errorf pigment content estimation by these spectral band combina-ions were calculated and used to identify spectral regions whereeflectance was most sensitive to the pigment of interest.

The validity of the bands included in the NN and PLS modelsas also investigated using the uninformative variable elimination

LS (UVE PLS) approach (Centner et al., 1996; Cai et al., 2008). Thisethod is based on a reliability measure c, which is computed using

he PLS regression coefficients for each wavelength:

wl =−bwl

std(bwl)

here−bwl is the mean regression coefficient, and std(bwl) is the

tandard deviation of the regression coefficient vector bwl . A lowbsolute value of the reliability parameter c indicates a less infor-ative band and such un-informative bands are removed from theodel (Centner et al., 1996; Cai et al., 2008).

The rededge chlorophyll index, CIrededge, (Gitelson et al., 2003)nd the rededge carotenoid reflectance index, CRIrededge (Gitelsont al., 2002, 2006) were used for Chl and Car estimation, respec-ively. Both VIs showed high accuracy for estimating pigment

ontents in a number of species (Richardson et al., 2002; Ustin et al.,009; Feret et al., 2011):

Ired edge = (�770/�710) − 1

ervation and Geoinformation 38 (2015) 251–260 253

CIred edge = [(�510)−1 − (�730)−1)] × �770

where �i is reflectance of the spectral band (20 nm width) centeredat the wavelength i.

Regardless of the model type, the calibration was performedusing 50% randomly chosen spectra. For the NN and PLS modelseach combination of three (for Chl content) or four (for Car con-tent) bands was tested 500 times, each time with different randomcalibration and validation sets. The errors of pigment estimationreported were obtained by averaging the validation errors obtainedin these 500 trials.

The vegetation indices were fitted to the corresponding pig-ment contents (Chl and Car) with 1st and 2nd order polynomials.The indices were validated with the remaining 50% of the spectra,and the errors of pigment estimation were obtained by averagingthe error measures of the 500 trials. Throughout the paper threemeasures of accuracy are used to describe the performance of themodels: the coefficient of variation, mean normalized bias, andnormalized mean bias (Simon et al., 2012).

Results and discussion

Model Calibration

The PLS and NN models for Chl estimation were based on 9reflectance measurements, three measurements for each band (e.g.,400, 410 and 420 nm for the band centered at 410 nm) and themodels for Car were based on 12 reflectance measurements.

Chlorophyll content

Relationships between the models (CIrededge, NN and PLS) andChl content were very close (Table 2). For all models examined,the coefficient of determination (R2) of the linear model vs. Chl washigher than 0.95. Due to very different ranges of pigment content inthe three species, the coefficient of variation (CV) was also used tocharacterize accuracy. Second order polynomial relationship Chl vs.CIrededge yielded a CV slightly higher than the linear model (11.3 vs.13.5%). CV varied marginally between models (9.8% through 13.5%)with the closest calibration relationship for NN with CV = 9.8%. Thelocations of the spectral bands (in nm) retained in the NN and PLSmodels for Chl estimation are:

460 − 480; 530 − 550; 730 − 750 nm (1)

Carotenoids content

Relationships between NN and PLS models, rededge carotenoidreflectance index, CRIrededge, and Car content were also very closefor each species (Table 2). However, while CRIrededge vs. Car rela-tionships for maple and chestnut had very similar slopes, the slopeof the relationship for beech was much lower than for maple andchestnut, thus the CRIrededge vs. Car relationship for all three speciestaken together was essentially not linear with CV above 23%. For allthree species taken together, NN vs. Car and PLS vs. Car relation-ships were much closer than between CRIrededge and Car with R2

above 0.93 and 0.91, respectively. As for Chl the lowest CV for allspecies taken together was achieved by NN (11.7%) followed by PLS(14%).

The locations of spectral bands retained in the NN model for Carestimation are:

490 − 510; 550 − 570; 610 − 630; 730 − 750nm (2)

254 O. Kira et al. / International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260

Table 2Descriptive statistics of the relationships between models examined, neural network, PLS, vegetation indices CIrededge for Chl and CRIrededge for Car and total leaf chlorophylland carotenoids content for calibration data set. CV is coefficient of variation in percent, R2 is coefficient of determination.

VI linear VI 2nd degree polynomial NN PLS

Chlorophyll CV R2 CV R2 CV R2 CV R2

Maple 14.6 0.96 12.4 0.96 11.6 0.97 12.6 0.96Chestnut 26.1 0.97 26.5 0.97 23.8 0.97 25.9 0.97Beech 10.1 0.97 7.3 0.98 5.1 0.99 6.2 0.99Total 13.5 0.96 11.3 0.97 9.8 0.98 10.9 0.97

CarotenoidsMaple 29.6 0.86 28.3 0.86 16.2 0.8 15.8 0.84Chestnut 22.5 0.89 20.8 0.87 10.2 0.93 13.7 0.85Beech 21 0.85 20.2 0.83 9.1 0.93 12.4 0.87Total 24.4 0.72 23.3 0.74 11.7 0.93 14 0.91

Table 3Descriptive statistics of the relationships between pigment content estimated by the models, NN, PLS and VI, CI rededge for Chl and CRI rededge for Car, and leaf chlorophylland carotenoids contents. CV is coefficient of variation, MNB is mean normalized bias, NMB is normalized mean bias; all three measures are in percent.

VI linear VI 2nddegree polynomial

CV R2 MNB NMB CV R2 MNB NMB

Chlorophyll Maple 15.2 0.96 59.5 −4.5 13.1 0.96 −4.7 −2.2Chestnut 27.6 0.96 69.6 19 28.1 0.97 3.3 13.1Beech 11 0.96 1.5 0.3 8 0.98 1.4 −0.9Three species 14.3 0.96 38.9 −1.1 12.1 0.97 −0.8 −0.1

NN PLSCV R2 MNB NMB CV R2 MNB NMB

Chlorophyll Maple 14 0.95 27.2 −1.4 14.5 0.95 −7.8 −2.3Chestnut 27.7 0.97 −14 9 29.1 0.96 −17.1 10.4Beech 6.5 0.98 0.2 −0.6 7.3 0.98 0.9 −0.5Three species 11.8 0.97 8.9 −0.3 12.5 0.97 −6.2 0.6

VI linear VI 2nd degree polynomialCV R2 MNB NMB CV R2 MNB NMB

Carotenoids Maple 30.7 0.85 18.1 21.2 30 0.85 20.6 21.1Chestnut 23.7 0.89 −3.5 2 22.7 0.86 0.9 4Beech 22 0.84 −9.8 −14.6 21.4 0.81 −10.8 −14.8Three species 25.5 0.69 3.3 1.5 24.8 0.7 4.7 0.9

NN PLSCV R2 MNB NMB CV R2 MNB NMB

Carotenoids Maple 20.1 0.71 1.4 0.6 19 0.8 0.2 2.2.6

.7

e

4

M

apz

TS(

Chestnut 13.8 0.87 1Beech 12.9 0.85 2Three species 15.6 0.88 1

The locations of spectral bands retained in the PLS model for Carstimation are:

70 − 490; 550 − 570; 610 − 630; 740 − 760nm (3)

odel validation

All three models were found to be estimating Chl content

ccurately (Table 3; Fig. 1). The best model was the 2nd degreeolynomial model for CIrededge with CV below 12.1% and almostero bias. NN as well as PLS models were also very accurate with

able 4pecifications of the seven spectral bands (B1 B7) of the Multi Spectral InstrumentMSI) aboard the Sentinel-2 satellite.

Spectral band Center wavelength(nm)

Band width(nm)

Spatial resolution(m)

B1 443 20 60B2 490 65 10B3 560 35 10B4 665 30 10B5 705 15 20B6 740 15 20B7 783 20 20

1.9 17.4 0.76 3.7 1.9−1.9 14.6 0.82 −13.6 −13.9

1.4 16.7 0.86 −4.5 17.7

CV below 11.8% for NN with 8.9% positive MNS and 12.5% for PLSwith 9% negative MNB.

Two models, PLS and NN, estimated Car content very accurately(Table 4, Fig. 2). The best results were achieved using PLS (CV below14%) with 2% MNB followed by NN with CV = 15.9% and MNB = 1.7%.The 2nd order polynomial CRIrededge model was far behind PLS andNN with CV above 24.8% and MNB = 4.7%.

Most significant spectral bands for pigment estimation

In the critique of stepwise multiple linear regression for theextraction of leaf biochemistry information from leaf reflectancedata, Grossman et al. (1996) questioned the meaningfulness of highdetermination coefficients obtained by regressions for several rea-sons:

- the lack of correspondence among the bands selected by stepwisemultiple linear regression between datasets and those reported

in other studies,

- the dependence of band selection on the data used, and- the inability of known absorption bands to explain the chemical

variation in their datasets.

O. Kira et al. / International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260 255

0

100

200

300

400

500

600

700

800

0 100 200 300 400 500 600 700 800

Chl

orop

hyll

pred

icte

d, m

g m

-2

Chlorophyll measured, mg cm-2

maple

chestnut

beech

y = 0.961x + 11. 586R2 = 0.96, STD = 40 mg m-2

A

0

100

200

300

400

500

600

700

800

0 100 20 0 30 0 400 500 600 700 800

Chl

orop

hyll

pred

icte

d, m

g m

-2

Chloroph yll measure d, mg cm-2

maple

chestnu t

beech

y = 0.970x + 8.357R2 = 0.97, STD = 34 mg m-2

B

0

100

200

300

400

500

600

700

800

0 10 0 200 300 400 500 60 0 700 80 0

Chl

orop

hyll

pred

icte

d, m

g m

-2

Chl oroph yll measure d, mg cm-2

maple

chestnu t

beech

y = 0.971x + 8.3 36R2 = 0.97, STD = 33 mg m-2

C

0

100

200

300

400

500

600

700

800

0 100 200 30 0 400 500 600 700 80 0

Chl

orop

hyll

pred

icte

d, m

g m

-2

Chloroph yll measure d, mg cm-2

maple

chestnu t

beech

y = 0.973x + 7.249R2 = 0.97, STD = 35 mg m-2

D

F ll indeC

bwaiefiabf

wtemwiCaviN

i(e

ig. 1. Chlorophyll content estimated by the models, (A) and (B) rededge chlorophyhl vs. CIrededge relationship; (B) 2nd order polynomial Chl vs. CIrededge relationship.

This is an important issue and we tried to answer these questionsy comparing the spectral bands retained in the NN and PLS modelsith those retained by two other methods that attempt to evalu-

te how informative the spectral bands are. In the first method,n order to find the most significant spectral bands for pigmentstimation, one band (among three for Chl and four for Car) wasxed and the positions of the others was varied. Spectra of aver-ge error of pigment estimation by the models using these spectraland combinations were calculated and presented in Fig. 3A and Bor Chl and Car, respectively.

For Chl estimation, minimal CV (below 14%) was achieved atavelength 730–760 nm (Fig. 3A). Local minimum of error was in

he green range around 560 nm and in the range of so-called greendge around 530 nm. However, errors of Chl estimation at theseinima were much higher (22–23%) than in rededge/NIR region atavelength 730–760 nm where global minimum was found. Max-

mal error was in the range around 680 nm, location of in situ redhl-a absorption band, and in the blue range of the spectrum wherell chlorophylls and carotenoids absorb. Importantly, CV spectrumaried widely – from 13.3% at 730 nm to 28% at 680 nm. Anothermportant finding is that spectral behavior of CV for both models,N and PLS, coincided precisely (Fig. 3A).

Minimal error of Car estimation, below 20%, was achievedn the rededge/NIR range 740–760 nm (Fig. 3B). Other minima22–22.5%) located in the green range around 560 nm and greendge at 520 nm. As for Chl, maximal errors of Car estimation were

x CIrededge, (C) NN and (D) PLS, plotted vs. measured chlorophyll content. (A) Linear

detected as blue or red bands were fixed in the four band combi-nation.

The results of the UVE PLS approach are presented in Fig. 4. Itis worth to mention that Centner et al. (1996) cautioned that their“approach is not a band selection in the sense that one tries to findthe best small subset of variables for fitting a model, but the elim-ination of those variables that are useless”. So, we considered themagnitude of the reliability parameter as an indicator of the infor-mation contained in the spectral bands and compared the mostinformative bands with the spectral bands used by the other models(NN, PLS models and VIs, CIrededge and CRIrededge).

The green spectral band around 560 nm retained in the NN andPLS models for Chl estimation (red areas in Fig. 4A) coincided withhighest values of the reliability parameter. In this spectral regionreflectance is governed by Chl content hyperbolically decreasingwith increase of Chl (Gitelson and Merzlyak, 1996). This spectralrange was widely used for Chl estimation due to high sensitivity ofreflectance to Chl content in a wide range of its change from slightlygreen to dark green leaves (Gausman et al., 1969; Fukshansky et al.,1993; Buschmann and Nagel, 1993; Gitelson and Merzlyak, 1994;Merzlyak and Gitelson, 1995).

Another maximum of the reliability parameter was found in

the long wave end of the rededge region between 730 and 750 nmwhere two factors govern reflectance – Chl absorption that may bestill significant in leaves with Chl content above 400 mg m−2 (greento dark green leaves), and leaf structure and thickness affecting

256 O. Kira et al. / International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260

0

20

40

60

80

100

120

140

0 20 40 60 80 10 0 12 0 140

Car

oten

oids

pre

dict

ed, m

g m

-2

Carotenoids measured , mg cm-2

maple

chestnut

beech

y = 0.691 x + 19 .22 3R2 = 0.69, STD = 16 mg m-2

A

0

20

40

60

80

100

120

140

0 20 40 60 80 10 0 120 140

Car

oten

oids

pre

dict

ed, m

g m

-2

Caroten oids measure d, mg cm-2

maple

chestnu t

beech

y = 0.707 x + 18 .38 8R2 = 0.70, STD = 16 mg m-2

B

0

20

40

60

80

100

120

140

0 20 40 60 80 10 0 12 0 140

Car

oten

oids

pre

dict

ed, m

g m

-2

Carotenoids measured , mg cm-2

maple

chestnut

beech

y = 0. 896x + 6.3 00R2 = 0.88, STD = 10 mg m-2

C

0

20

40

60

80

100

120

140

0 20 40 60 80 100 120 140

Car

oten

oids

pre

dict

ed, m

g m

-2

Carotenoids measured , mg cm-2

maple

chestnut

beech

y = 0.87 0x + 7.99 3R2 = 0.86, STD = 11 mg m-2

D

Fig. 2. Carotenoids content estimated by the models, (A) and (B) rededge carotenoid reflectance index, CRIrededge, (C) NN and (D) PLS plotted vs. measured carotenoids content.(A) Linear Car vs. CIrededge relationship; (B) 2nd order polynomial Car vs. CIrededge relationship.

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

760720680640600560520480440400

Coe

ffic

ient

of

vari

atio

n, %

Wavelength, nm

NN

PLS

A

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

760720680640600560520480440400

Coe

ffic

ient

of

vari

atio

n, %

Wavelength, nm

NN

PLS

B

Fig. 3. Spectra of average error of (A) chlorophyll and (B) carotenoid content estimation by NN and PLS models. One band (among three for Chl and four for Car) was fixedand the positions of others varied. The number of combinations was 496 for Chl and 4060 for Car. The average error of pigment content estimation by models with thesespectral band combinations shows spectral regions where reflectance was most sensitive to the pigment of interest.

O. Kira et al. / International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260 257

Fig. 4. Absolute value of reliability parameter calculated using uninformative variable elimination PLS (UVE PLS) versus wavelength for (A) chlorophyll content and (B)c ands

( e optio sion o

r2ba2e

tiaCtrpma

eoo(

arotenoids content estimation. The gray areas at the top correspond to Sentinel-2 bTable 5). The red areas indicate the 20 nm wide spectral bands which were found to bf the references to color in this figure legend, the reader is referred to the web ver

eflectance in the NIR range (Fukshansky et al., 1993; Feret et al.,011; le Maire et al., 2004; Gitelson, 2011). In previous studies, thisand was used for accurate estimating foliar and total canopy Chlnd nitrogen content using rededge Chl index (Clevers and Kooistra,012; Clevers and Gitelson, 2013; Inoue et al., 2012; Schlemmerst al., 2013).

Despite the very low magnitude of the reliability parameter inhe blue region (Fig. 4A), the band around 480 nm was retainedn the NN and PLS models for Chl estimation (1). In this regionbsorption is saturated strongly and sensitivity of reflectance tohl content is minimal, which is indicated by zero magnitude ofhe reliability parameter. However, it was shown that the blueange is suitable for reference reflectance used in VIs for eliminatingartially random variability of reflectance due to uncertainties ofeasurements as well as differences in leaf surface structure (Sims

nd Gamon, 2002).Overall we can conclude that the reliability parameter for Chl

stimation indeed provides a reliable indication of the usefulnessf the spectral region: its maxima coincided with the positionsf spectral bands used by the green and rededge Chl indicesGitelson et al., 2003, 2005). The other bands used in these indices

(Table 4), which were found optimal for pigment estimation by NN and PLS modelsmal for pigment content estimation by NN and PLS (Eqs. (1)–(3)). (For interpretationf this article.)

(beyond 770 nm and either 540–560 nm in CIgreen or 690–730 nmin CIrededge) also correspond to regions in which the magnitude ofthe reliability parameter is substantial.

All four spectral bands retained in the NN and PLS models for Carcontent estimation, (2) and (3), coincided with highest values of thereliability parameter (Fig. 4B) as well as with spectral bands of therededge and green CRI (Gitelson et al., 2002, 2006). The first bandretained in the models was located around 510 nm where maximalsensitivity of reflectance to Car content was found (Gitelson et al.,2002). In this region reflectance is governed by both Chl and Carcontent and this band was used in both CRIgreen and CRIrededge.

The second band retained in the models was located in the greenrange where the magnitude of the reliability parameter is maxi-mal. This band was used in CRIgreen for subtraction of Chl effectfrom reflectance around 510 nm. This waveband is located quitefar from main absorption bands of Chl absorption to avoid satura-tion of reflectance at moderate to high Chl values but close enough

to keep reflectance sensitive to Chl content.

The third and fourth spectral bands retained in the NN andPLS models were located around 630 and 740 nm. Both bands arelocated quite far from the main absorption bands of Chl and Car

258 O. Kira et al. / International Journal of Applied Earth Observation and Geoinformation 38 (2015) 251–260

Table 5Coefficient of variation (in percent) of chlorophyll and carotenoids estimation by neural network (NN) and partial least squares (PLS) regression with simulated expectedspectral response of the Multi Spectral Instrument (MSI) aboard the Sentinel-2 satellite. The spectral bands of MSI are given in Table 4.

Bands CV Bands CV

NN PLS NN PLS

Chlorophyll B1 B7 13.6 16.0 Carotenoids B1 B7 19.9 21.0B4 B7 13.6 13.6 B4 B7 21.1 21.1B5 B7 13.6 13.6 B5 B7 21.0 21.0B6 B7 14.1 14.1 B6 B7 20.9 20.9B5 B6 26.9 26.8 B5 B6 27.9 28.0B4 B6 20.9 20.9 B4 B6 25.9 25.9

srbttafh

caacsia2nacss(rx(

P(rrdsis

l(swatobtitaeub

B2, B4, B6 B7 13.7 13.7

B3, B6, B7 13.6 13.6B3, B5, B7 21.5 21.4

o absorption is not saturated at moderate to high Chl content buteflectance in these spectral regions is sensitive to Chl content. Theand centered at 630 nm has an additional advantage – absorp-ion by both chlorophyll-a and chlorophyll-b at 630 nm are almosthe same. Thus, in contrast to the 740 nm band where only Chl-absorbs, the 630 nm band is accounted not only for Chl-a but alsoor Chl-b absorption. To the best of our knowledge, the 630 nm bandas never been used in any estimation of either Chl or Car contents.

The location of the spectral bands retained in the models did notorrespond to main absorption bands of pigment of interest, Chlnd Car. They coincided with the location of spectral bands wherebsorption by Chl remained strong enough to be sensitive to Chlontent but far enough from the main Chl absorption bands to avoidaturation (540 and 730 nm for Chl models and 560, 630 and 740 nmn Car models). Chl models also used a band in the blue range toccount for differences in leaf surface reflectance (Sims and Gamon,002) and also as a reference, presumably subtracting randomoise due to uncertainties of reflectance measurements. Remark-bly, for Car estimation the NN and PLS models retained bandsentered at 490–510 nm and 470–490 nm, respectively, which arepectral region where reflectance was found to be maximally sen-itive to Car absorption while being affected also by Chl absorptionChappelle at al., 1992; Blackburn, 1998; Gitelson et al., 2002). Thisegion is also close to the range where reflectance is sensitive toanthophyll cycle and used in photochemical reflectance indexGamon et al., 1992).

The repeatability of the wavelength selection for the NN andLS models is remarkable, they almost completely coincided (Eqs.2) and (3); Figs. 3 and 4). It is also worth noting that consistentesults of Chl and Car estimation by NN and PLS were achieved usingeflectance without any spectral transformation (e.g., log(1/R), firsterivative, second derivative) (Grossman et al., 1996). The bandelection was not dependent on the data used; the bands retainedn the NN and PLS models agreed well with those reported in othertudies and known to explain the chemical variation in our datasets.

The hyperspectral reflectance spectra were resampled to simu-ate the expected spectral response of the Multi Spectral InstrumentMSI) on the Sentinel-2 satellite (Table 4). Sentinel-2 bands wereimulated by calculating the average reflectance over the bandidth of the respective Sentinel-2 bands. This approximation was

pplied since the spectral response functions of the Sentinel-2 spec-ral bands are close to rectangular (Drusch et al., 2012). Accuracyf pigment content estimation by VIs, NN and PLS with spectralands of Sentinel-2 was tested. CIrededge with spectral bands cen-ered at 705 and 775 nm (Table 4) was able to estimate Chl contentn all three species taken together with CV = 13.1%, while with spec-ral bands at 740 and 775 nm CV was even lower −12.6%. The CV

chieved by the NN and PLS models presented in Table 5. Minimalrror of Chl estimation by both NN and PLS models was achievedsing all seven spectral bands (CV = 13.53%), but using only threeands 540–580, 732.5–747.5 and 770–780 nm allowed for accurate

B2, B5, B6 B7 20.5 20.4B2, B4, B6 B7 19.6 19.6

estimation of Chl with CV = 13.63%. The CV achieved by NN and PLSusing the Sentinel MSI spectral bands was only 1–1.5% higher thanusing optimal spectral bands (Table 5).

CRIrededge employing bands centered at 490 nm, 705 nm, and783 nm was able to estimate Car content in all three species takentogether with CV below 27%. CV was a little bit higher (28.5%) whenthe band at 705 nm was replaced by the band at 740 nm. Car esti-mation by NN and PLS was more accurate than that by CRIrededge(Table 5). The highest accuracy (CV = 19.6%) was achieved using fourspectral bands (B2, B4, B6, B7) in blue, red, rededge and NIR rangesof the spectrum. CV of Car estimation using MSI bands was about3–4% higher than using 20 nm wide optimal bands. This is likelydue to the use of the band B2 positioned in the green edge regionbetween 460 and 525 nm. The width of this band does not cor-respond to the required 10–15 nm width of the band positionedat 510 nm where maximal sensitivity of reflectance to Car contentwas found (Chappelle et al., 1992; Gitelson et al., 2002, 2006).

While green and rededge CI were tested at close range at canopylevel (Gitelson et al., 2005) as well as using TM Landsat data (Wuet al., 2010), Car indices were not tested at canopy level. It is nec-essary to examine the presented techniques also at other scales.For now it remains unclear whether the found linear relationshipsbetween the models and pigment content also hold on a coarserspatial scale. It has to be examined if the proposed techniques areable to estimate Car content at canopy scale using airborne andsatellite data which is typically influenced by atmosphere, bidi-rectional reflectance distribution function effects, canopy shadowsand soil background.

Another type of index for Chl estimation based on the MERISred-edge bands is the MERIS terrestrial chlorophyll index, MTCI(Dash and Curran, 2004). This index has been applied success-fully for many applications. As heritage of MERIS, this MTCI is thebasis of one of the Level 2B main terrestrial products of Sentinel-2and Sentinel-3, called the OLCI Terrestrial Chlorophyll Index (OTCI)(Vuolo et al., 2012). Therefore, this index will also get explicit atten-tion when three techniques presented in this paper would be testedat canopy level.

Conclusions

Over the last decade, technological developments have made itpossible to quickly and nondestructively assess, in situ, the chloro-phyll and carotenoids status of plants. The focus of this studywas to identify informative spectral bands of three types of mod-els, vegetation indices, neural network and partial least squaresregression, and test their performance for estimating pigmentscontents in three unrelated tree species. All three models were

found to provide accurate estimations of foliar Chl content forthree tree species combined together. The chlorophyll index usingonly two spectral bands in the rededge and NIR may be recom-mended for Chl estimation. NN and PLS with four spectral bands

th Obs

weaircceShfa

A

C

R

B

B

B

B

B

B

C

C

C

C

C

C

C

C

D

D

D

F

F

F

O. Kira et al. / International Journal of Applied Ear

ere more accurate than rededge carotenoid reflectance index forstimating carotenoids content; the NN model showed the highestccuracy. All techniques tested were not species-specific, allow-ng for estimating pigment content in different species withoute-parameterization of the model. All three techniques performedonsistently well and yielded accurate estimations of pigmentontent when spectral bands were simulated in accord with thexpected spectral response of the Multi Spectral Instrument on theentinel-2 satellite. We are convinced that the analyses presentedere will facilitate the quantitative non-destructive estimation of

oliar chlorophyll and carotenoids and will therefore provide a valu-ble and fruitful addition to these efforts.

cknowledgement

This research was supported by International Incoming Marieurie Fellowship to AG.

eferences

aret, F., Houlès, V., Guérif, M., 2007. Quantification of plant stress using remotesensing observations and crop models: the case of nitrogen management. J.Exp. Bot. 58, 869–880.

lackburn, G.A., 2007a. Hyperspectral remote sensing of plant pigments. J. Exp.Bot. 58, 855–867.

lackburn, G.A., 2007b. Wavelet decomposition of hyperspectral data: a novelapproach to quantifying pigment concentrations in vegetation. Int. J. RemoteSens. 28, 2831–2855.

lackburn, G.A., 1998. Quantifying chlorophylls and carotenoids at leaf and canopyscales: an evaluation of some hyperspectral approaches. Remote Sens. Environ.66, 273–285.

rereton, R., 2005. Chemometrics, Data Analysis for the Laboratory and ChemicalPlant. John Wiley & Sons Ltd.

uschmann, C., Nagel, E., 1993. In vivo spectroscopy and internal optics of leaves asbasis for remote sensing of vegetation. Int. J. Remote Sens. 14, 711–722.

ai, W., Yankun, Y., Shao, X., 2008. A variable selection method based onuninformative variable elimination for multivariate calibration ofnear-infrared spectra. Chemom. Intell. Lab. 90, 188–194.

entner, V., Massart, D.-L., de Noord, O.E., de Jong, S., Vandeginste, B.M., Sterna, C.,1996. Elimination of uninformative variables for multivariate calibration. Anal.Chem. 68, 3851–3858.

happelle, E.W., Kim, M.S., McMurtrey III, J.E., 1992. Ratio analysis of reflectancespectra (RARS): an algorithm for the remote estimation of the concentrationsof chlorophyll a, chlorophyll b, and carotenoids in soybean leaves. RemoteSens. Environ. 39, 239–247.

hen, L., Huang, J.F., Wang, F.M., Tang, Y.L., 2007. Comparison between backpropagation neural network and regression models for estimation of pigmentcontent in rice leaves and panicles using hyperspectral data. Int. J. RemoteSens. 28, 3457–3478.

levers, J.G.P.W., Kooistra, L., 2012. Using hyperspectral remote sensing data forretrieving canopy chlorophyll and nitrogen content. IEEE J. Selected Topics inAppl. Earth Obs. Remote Sens. 5, 574–583.

levers, J.G.P.W., Gitelson, A.A., 2013. Remote estimation of crop and grasschlorophyll and nitrogen content using red-edge bands on Sentinel-2 and -3.Int. J. Appl. Earth Obs. Geoinf. 23, 344–351.

oops, N., Drury, S., Smith, M.L., Martin, M., Ollinger, S., 2002. Comparison of greenleaf eucalypt spectra using spectral de- composition. Aust. J. Bot. 50, 567–576.

urran, P.J., Dungan, J.L., Gholz, H.L., 1990. Exploring the relationship betweenreflectance red-edge and chlorophyll content in slash pine. Tree Physiol. 7,33–48.

emmig-Adams, B., Adams, W.W., 1996. The role of xanthophyll cycle in theprotection of photosynthesis. Trends Plant Sci. 1, 20–26.

ash, J., Curran, P.J., 2004. The MERIS terrestrial chlorophyll index. Int. J. RemoteSens. 25, 5403–5413.

rusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B.,Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F.,Bargellini, P., 2012. Sentinel-2: ESA’s optical high-resolution mission for GMESoperational services. Remote Sens. Environ. 120, 25–36.

eret, J.B., Franc ois, C., Asner, G.P., Gitelson, A.A., Martin, R.E., Bidel, L.P.R., Ustin,S.L., le Maire, G., Jacquemoud, S., 2008. PROSPECT-4 and 5: advances in the leafoptical properties model separating photosynthetic pigments. Remote Sens.Environ. 112, 3030–3043.

éret, J.-B., Franc ois, C., Gitelson, A.A., Barry, K.M., Panigada, C., Richardson, A.D.,Jacquemoud, S., 2011. Optimizing spectral indices and chemometric analysis of

leaf chemical properties using radiative transfer modeling. Remote Sens.Environ. 115, 2742–2750.

ilella, I., Serrano, L., Serra, J., Penuelas, J., 1995. Evaluating wheat nitrogen statuswith canopy reflectance indices and discriminant analysis. Crop Sci. 35,1400–1405.

ervation and Geoinformation 38 (2015) 251–260 259

Gamon, J.A., Penuelas, J., Field, C.B., 1992. A narrow-waveband spectral index thattracks diurnal changes in photosynthetic efficiency. Remote Sens. Environ. 41,35–44.

Fukshansky, L.A., Remisowsky, A.M., McClendon, J., Ritterbusch Richter, A., Mohr,H., 1993. Absorption spectra of leaves corrected for scattering anddistributional error: a radiative transfer and absorption statistics treatment.Photochem. Photobiol. 57, 538–555.

Gamon, J.A., Qiu, H., 1999. Ecological applications of remote sensing at multiplescales. In: Pugnaire, F.I., Valladares, F. (Eds.), Handbook of functional plantecology. Marcel Dekker, New York, pp. 805–846.

Gausman, H.W., Allen, W.A., Cardenas, R., 1969. Reflectance of cotton leaves andtheir structure. Remote Sens. Environ. 1, 19–22.

Gitelson, A.A., 2011. Non-destructive Estimation of Foliar Pigment (chlorophylls,Carotenoids and Anthocyanins) Contents: Espousing a Semi-analyticalThree-band Model. Chapter 6 in Hyperspectral Remote Sensing of Vegetation.In: Thenkabail, P.S., Lyon, J.G., Huete, A. (Eds.). Taylor and Francis,pp. 141–165.

Gitelson, A., Merzlyak, M., 1994. Quantitative estimation of chlorophyll-a usingreflectance spectra: experiments with autumn chestnut and maple leaves. J.Photochem. Photobiol. B: Biol. 22, 247–252.

Gitelson, A., Merzlyak, M., 1996. Signature analysis of leaf reflectance spectra:algorithm development for remote sensing of chlorophyll. J. Plant Physiol. 148,495–500.

Gitelson, A.A., Merzlyak, M.N., Chivkunova, O.B., 2001. Optical properties andnon-destructive estimation of anthocyanin content in plant leaves.Photochem. Photobiol. 74 (1), 38–45.

Gitelson, A.A., Zur, Y., Chivkunova, O.B., Merzlyak, M.N., 2002. Assessing carotenoidcontent in plant leaves with reflectance spectroscopy. Photochem. Photobiol.75 (3), 272–281.

Gitelson, A.A., Gritz, U., Merzlyak, M.N., 2003. Relationships between leafchlorophyll content and spectral reflectance and algorithms fornon-destructive chlorophyll assessment in higher plant leaves. J. Plant Physiol.160, 271–282.

Gitelson, A.A., Vina, A., Rundquist, D.C., Ciganda, V., Arkebauer, T.J., 2005. Remoteestimation of canopy chlorophyll content in crops. Geophys. Res. Lett. 32,L08403, http://dx.doi.org/10.1029/2005GL022688.

Gitelson, A.A., Keydan, G.P., Merzlyak, M.N., 2006. Three-band model fornoninvasive estimation of chlorophyll, carotenoids, and anthocyanin contentsin higher plant leaves. Geophys. Res. Lett. 33, L11402,http://dx.doi.org/10.1029/2006GL026457.

Grossman, Y., Ustin, S., Jacquemoud, S., Sanderson, E., Schmuck, G., Verdebout, J.,1996. Critique of stepwise multiple linear regression for the extraction of leafbiochemistry information from leaf reflectance data. Remote Sens. Environ. 56,182–193.

Hatfield, J.L., Gitelson, A.A., Schepers, J.S., Walthall, C.L., 2008. Application ofspectral remote sensing for agronomic decisions. Agron. J. 100, S-117–S-131,http://dx.doi.org/10.2134/agronj2006.037c.

Haykin, S., 2005. Neural Networks, A Comprehensive Foundation, second edition.Pearson Education, Inc.

Jacquemoud, S., Baret, F., 1990. PROSPECT: a model of leaf optical properties.Remote Sens. Environ. 34, 75–91.

Inoue, Y., Sakaiy, E., Zhu, Y., Takahashi, W., 2012. Diagnostic mapping of canopynitrogen content in rice based on hyperspectral measurements. Remote Sens.Environ. 126, 210–221.

Kergoat, L., Lafont, S., Arneth, A., Le Dantec, V., Saugier, B., 2008. Nitrogen controlsplant canopy light-use-efficiency in temperate and boreal ecosystems. J.Geophysi. Res. 113 (G04017), 1 http://dx.doi.org/10.1029/2007JG000676

Larcher, W., 1995. Physiological Plant Ecology, 3rd edn. Springer, Berlin, Germany.le Maire, G., Francois, C., Dufrene, E., 2004. Towards universal broad leaf

chlorophyll indices using PROSPECT simulated database and hyperspectralreflectance measurements. Remote Sens. Environ. 89, 1–28.

Maier, S.W., Lüdeker, W., Günther, K.P., 1999. SLOP: a revised version of thestochastic model for leaf optical properties. Remote Sens. Environ. 68,273–280.

Merzlyak, M., Gitelson, A., 1995. Why and what for the leaves are yellow inautumn? On the interpretation of optical spectra of senescing leaves (Acerplatanoides L.). J. Plant Physiol. 145 (3), 315–320.

Moran, J.A., Mitchell, A.K., Goodmanson, G., Stockburger, K.A., 2000. Differentiationamong effects of nitrogen fertilization treatments on conifer seedlings by foliarreflectance: a comparison of methods. Tree Physiol. 20, 1113–1120.

O’Neill, A.L., Kupiec, J.A., Curran, P.J., 2002. Biochemical and reflectance variationthroughout a Sitka spruce canopy. Remote Sens. Environ. 80, 134–142.

Osborne, S.L., Schepers, J.S., Francis, D.D., Schlemmer, M.R., 2002. Use of spectralradiance to estimate in-season biomass and grain yield in nitrogen- andwater-stressed corn. Crop Sci. 42, 165–171.

Penuelas, J., Filella, I., 1998. Visible and near-infrared reflectance techniques fordiagnosing plant physiological status. Trends Plant Sci. 3, 151–156.

Richardson, A.D., Duigan, S.P., Berlyn, G.P., 2002. An evaluation of noninvasivemethods to estimate foliar chlorophyll content. New Phytol. 153, 185–194.

Schlemmer, M., Gitelson, A.A., Schepers, J., Ferguson, R., Peng, Y., Shanahan, J.,Rundquist, D.C., 2013. Remote estimation of nitrogen and chlorophyll contents

in maize at leaf and canopy levels. Int. J. Appl. Earth Obs. Geoinf. 25,47–54.

Simon, H., Baker, K.R., Phillips, S., 2012. Compilation and interpretation ofphotochemical model performance statistics published between 2006 and2012. Atmos. Environ. 61, 124–139.

2 th Obs

S

T

U

60 O. Kira et al. / International Journal of Applied Ear

ims, D.A., Gamon, J.A., 2002. Relationships between leaf pigment content andspectral reflectance across a wide range of species, leaf structures anddevelopmental stages. Remote Sens. Environ. 81, 337–354.

umbo, S.D., Wagner, D.G., Heinemann, P.H., 2002. Hyperspectral- based neural

network for predicting chlorophyll status in corn. Trans. ASAE 45, 825–832.

stin, S.L., Gitelson, A.A., Jacquemoud, S., Schaepman, M., Asner, G.P., Gamon, J.A.,Zarco-Tejada, P., 2009. Retrieval of foliar information about plant pigmentsystems from high resolution spectroscopy. Remote Sens. Environ. 113,S67–S77, http://dx.doi.org/10.1016/j.rse.2008.10.019.

ervation and Geoinformation 38 (2015) 251–260

Vuolo, F., Dash, J., Curran, P.J., Lajas, D., Kwiatkowska, E., 2012. Methodologies anduncertainties in the use of the terrestrial chlorophyll index for the Sentinel-3mission. Remote Sens. 4, 1112–1133.

Wu, C., Wang, L., Niu, Z., Gao, S., Wu, M., 2010. Nondestructive estimation of

canopy chlorophyll content using hyperion and landsat/TM images. Int. J.Remote Sens. 31, 2159–2167.

Yao, H.B., Tian, L., 2003. A genetic algorithm-based selective principal componentanalysis (GA-SPCA) method for high-dimensional data feature extraction. IEEETrans. Geosci. Remote Sens. 41, 1469–1478.