MURDOCH RESEARCH REPOSITORY

This is the author’s final version of the work, as accepted for publication following peer review. The definitive version is available at

http://remotesensing.spiedigitallibrary.org/article.aspx?articleid=1475464&journalid=96

Evans, B., Lyons, T., Barber, P., Stone, C. and Hardy, G. (2012) Enhancing a eucalypt crown condition indicator driven by high

spatial and spectral resolution remote sensing imagery. Journal of Applied Remote Sensing, 6 (1).

http://researchrepository.murdoch.edu.au/12194/ © 2012 Society of Photo-Optical Instrumentation Engineers. One print or electronic copy may be made for personal use only. Systematic reproduction and distribution, duplication of any material in this paper for a fee or for commercial purposes, or modification of the content of the paper are prohibited.

Enhancing a eucalypt crown condition indicator drivenby high spatial and spectral resolution remote sensing

imagery

Bradley Evans,a,b Tom Lyons,b Paul Barber,b,c Christine Stone,d andGiles Hardyb

aMacquarie University, Department of Biological Sciences, University Avenue, Macquarie Park,New South Wales, 2113, Australia

bradley.evans@mq.edu.aubMurdoch University, Centre of Excellence for Climate Change, Woodland and Forest Health, 90

South Street, Murdoch, Western Australia, 6150, AustraliacArbor Carbon Environmental and Arborcultural Consultants P.O. BOX 1065 Willagee Central,

Western Australia, 6156, AustraliadForest Science Centre, NSW Department of Primary Industries, P.O. Box 100, Beecroft, NSW,

Australia, 2119

Abstract. Individual crown condition of Eucalyptus gomphocephala was assessed using twoclassification models to understand changes in forest health through space and time. Usinghigh resolution (0.5 m) digital multispectral imagery, predictor variables were derived fromtextural and spectral variance of all pixels inside the crown area. The results estimate crowncondition as a surrogate for tree health against the total crown health index. Crown conditionis derived from combining ground-based crown assessment techniques of density, transparency,dieback, and the regrowth of foliage. This object-based approach summarizes the pixel data intomean crown indices assigned to crown objects which became the carrier of information. Modelsperformed above expectations, with a significant weighted Cohen’s kappa (κ > 0.60 andp < 0.001) using 70% of available data. Using in situ data for model development, crowncondition was predicted forwards (2010) and backwards (2007) in time, capturing trends incrown condition and identifying decline in the healthiest between 2008 and 2010. Theresults confirm that combining spectral and textural information increased model sensitivityto small variations in crown condition. The methodology provides a cost-effective meansfor monitoring crown condition of this or other eucalypt species in native and plantationforests. © 2012 Society of Photo-Optical Instrumentation Engineers (SPIE). [DOI: 10.1117/1.JRS.6.XXXXXX]

Keywords: tree crown condition; forest health; high resolution; digital multispectral imagery;remote sensing; Eucalyptus gomphocephala; tuart; random forest.

Paper 12050 received Mar. 2, 2012; revised manuscript received Oct. 22, 2012; accepted forpublication Oct. 25, 2012.

1 Introduction

Quantitative indicators of tree crown condition across large areas are a management imperativeand critical for the assessment of ecosystem biodiversity, health, and function.1,2 Symptoms ofcrown decline manifest as a reduction in the abundance, distribution, and greenness of foliage(dieback) within the tree crown and an overall reduction of vigor and condition.3 This diebackcan ultimately lead to mortality despite eucalypts being well-adapted to extreme climatevariability.4 In recent years, a number of declines in eucalypt dominated ecosystems have beendocumented.5–8 Similar to the findings in other areas, Eucalyptus gomphocephala (tuart)declines in southwestern Australia have been linked to a range of causes, including soil bacterialcommunities,9 pathogens,6 and fire frequency and intensity.10–12

0091-3286/2012/$25.00 © 2012 SPIE

1

Journal of Applied Remote Sensing XXXXXX-1 Vol. 6, 2012

Ground-based, or in situ, forest condition assessments of the extent, condition, and distribu-tion of foliage through a tree crown are a common method for assessing changes in health overtime.13 Despite their operational usefulness, they are considered a subjective, semi-quantitativeobservation and given the size and remoteness of the forest resource, labor intensive and timeconsuming.14 As a result, regular monitoring of forest condition over time is difficult yet is amanagement imperative under changing climatic and land use pressures.

Remote sensing methods and imagery from space-borne or airborne platforms offer analternative and viable and cost-effective means of assessing forest condition over large areasin a way that is spatially explicit and repeatable through time.15–17 Land Monitor II, a Land-sat-based vegetation trend product has been demonstrated to be useful in monitoring vegetationchanges over time.18,19 However, the spatial resolution (30 m) of the Landsat Thematic Mapper(TM) sensor limits the capacity to detect individual tree crowns and in-crown foliage variability.Alternatively, airborne high spatial resolution remote sensing imagery (0.5 m) offers greaterflexibility in image acquisition, the capacity to detect individual crowns and crown architecture.When used together, the two resolutions enable the combination of broad area temporal analysisand limited area crown condition analysis.

At both resolutions, the spectral information is often used independently and combined as avegetation index (VI) to predict condition, such as the distribution and density of foliage. VIsummarize spectral information from specific regions of the electromagnetic spectrum known tobe sensitive to the interactions between light, leaf, and canopy structure and pigments (mostlychlorophyll).20–22 In high resolution data, where many pixels are captured across individualcrowns, the spectral characteristics of neighboring pixels was highly variable and as a result,statistics produced in the spatial domain may yield more reliable estimates of crown conditionthan spectral indicators alone.23 A number of techniques exist to measure the variation of spectraland spatial properties of remotely sensed images including texture and geometric measures.

Stone and Haywood24 demonstrated the use of a eucalypt crown condition index derivedfrom remote sensing observations over areas of eucalypt dieback in New South Wales. Theapproach utilized high spatial resolution digital imagery linked to five key indicators of canopydecline. Recently, it was demonstrated in Ref. 25 that a similar index could be used to predictcrown dieback in Western Australia from remote sensing imagery. The results indicated thatcrown condition could be accurately modeled using the spectral properties of the imagerywith significant coefficient of determination (R2) values developed from linear regression mod-els ranging from 0.54 to 0.72.

The objectives of this study were threefold. First, we evaluated the additional power of tex-ture and geometric-based indices to predict canopy condition across crowns exhibiting diebackin the Yalgorup National Park of Western Australia. Second, we applied an innovative modelingapproach, random forests (RF), a nonparametric, ensemble classifier, to assess the impact ofvarying the number of predicted canopy condition classes with the aim of developing accurateand robust indicators for forest managers.26,27 Lastly, with confidence in the developed models,we then applied them to images acquired before and after the in situ assessments to better under-stand how the condition of these crowns has changed over time.

2 Methods

2.1 Region of Interest

The Yalgorup National Park, in southwest, Western Australia, is situated 80 vkm south of thecapital, Perth (Fig. 1). The sites were located within the Yoongarillup, Karrakatta, and Cottesloevegetation complexes, based on their geological associations.28 The majority of the sites werefound on the Yoongarillup complex, comprised of E. gomphocephala(tuart), with a mid-storeyof Agonis flexuosa, Allocasuarina fraseriana, Banksia grandis, B. attenuata, B. littoralis, and anunderstorey of Acacia saligna, A. pulchella, Jacksonia sternbergiana, Melaleuca acerosa, andHibbertia hypericoides.29 Tuart is listed by the International Union for the Conservation ofNature as IUCN category II species and Yalgorup National Park is an area of high naturalbiodiversity value.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-2 Vol. 6, 2012

2.2 Crown Condition Assessment

The placement of crown condition plots within the national park was based on a stratification ofvegetation trend classes derived from a 15-year (1990 to 2005) time series analysis of LandsatTM imagery based on the methodology of Caccetta et al.18 A total of 20 sites were established in2007 and in June 2008 the crown condition of four randomly selected tuart crowns within each ofthe 20 sites was individually assessed by two forest health professionals.

Each crown was assessed for four crown condition indices, based on previous research of Barryet al.21 and recently discussed in Cai et al.9; Evans et al.25; and Horton et al.30 The condition indiceswere measured in 5% classes across a 0% to 100% scale and included (1) crown density (A),(2) foliage transparency31,32 (B), (3) crown dieback ratio (C), and (4) an epicormic index (D).33,34

For each crown assessment, two trained forest health professionals stood at right anglesto each other at a distance from the main stem greater than its height. Two assessmentswere assumed to be more robust than a single observation and were averaged for an overallvalue. The location of each individual crown was mapped using a differential GPS(dGPS) for location in the imagery. All four individual crown condition indices were combinedto produce the total crown health index (TCHI), which weights crown density (A), foliagetransparency (B), epicormic index (C), and crown dieback ratio (D) equally(i.e., TCHI ¼ Aþ Bþ CþD∕4).25

2.3 Image Specification and Acquisition

High spatial resolution airborne digital multispectral imagery (DMSI) was acquired bySpecTerra Services Pty. Ltd. (Perth, Western Australia) in May/June 2007, 2008, and 2010

Fig. 1 Map of the study area overlaid with a true color version of the DMSI for 2008 and the studysites (irregularly numbered).

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-3 Vol. 6, 2012

under clear conditions (Table 1). During acquisition, all DMSI was georeferenced based on GPSand subsequent post-flight processing of the data included band to band registration and spectro-radiometric correction for bi-directional reflectance distribution function (BRDF) variations. Allimagery was acquired at 0.5 m resolution from an altitude of 2000 m. Once collected, allthe imagery was mosaicked, orthorectified, and pixel matched to correct for the sun angledifferences of each overpass (Table 1).

The DMSI sensor acquires 12-bit digital number (DN) data simultaneously in four narrowbands (20 nm full width half maximum) spectral channels in the visible and near-infrared regionsof the electromagnetic spectrum using filters centered at 450, 550, 675, and 780 nm coveringblue, green, red and near infrared regions of the spectrum, respectively.

To ensure the relative radiometric consistency of the imagery over the three time periods a“like-value” calibration was undertaken using methods described by Furby et al.35 The methodnormalizes each scene to a “base year” standard using a set of objects identifiable on all imagesand deemed to be invariant spectral targets (such as roads, deep water, exposed soil, and sand).Given that the in situ data were collected in 2008, this was chosen as the base year. For eachinvariant target a mean spectral value in all bands was extracted and matched using a simpleregression approach which was then subsequently applied to the 2007 and 2010 images.

Visual inspection of the crown dGPS delineations on the imagery confirmed that the crownswere all well located. In a small number of cases crown locations were aligned to better fit theimagery. The in situ crown condition assessment data were collated and spatially linked to thecrowns in the imagery and crown perimeters formed a tree crown mask. Spectral values withineach tree crown were then extracted for calculation of the indices.

2.4 Index Generation

2.4.1 Textural, spectral, and spatial indices

A total of 125 indices were selected based on (1) the available spectra, (2) literary precedence,and (3) the open source availability of the algorithms used for the gray level covariance matrixtextures (GLCM) and gray level difference vector textural indices (GLDV). The GRASSr.texture package was used to generate the textures.36,37

Indices fit into three broad categories of spectral (Table 2), textural (Table 3), and geometric(the area displaced by the crown by counting the number of pixels inside the delineated crownarea). The VI used in this study (Table 2) fall into two broad categories (1) narrow-band versionsof the normalized difference vegetation index (NDVI) and (2) simple ratios.

Textural indices were calculated from the covariance between neighboring gray levels in animage, resulting in a pixel score of relative positional differences. Rectilinear polygons wereformed around each object with a one pixel buffer to perform the GLCM and GLDV to ensureall pixels in the crown were used. Each textural index was calculated for each object in the fourcardinal directions. Each resulting textural index was then cropped and all four directions andaveraged by band.

Once the individual indices were computed, a mean for each was calculated for the 80 tuartcrowns, and this dataset formed the basis of the modeling described below.

Figure 2 provides an example of a true color DMSI image with four tree crown delineationsoverlaid and an example of two texture measures for an individual crown. The figure shows themarked variation within each crown, as well as the variation through time.

Table 1 Details of the multi-year acquisition flights made by SpecTerra Services.

Date Solar angle

May 08, 2007 39.18 deg to 40.08 deg

June 06, 2008 34.01 deg to 34.60 deg

June 02, 2010 34.74 deg to 35.03 deg

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-4 Vol. 6, 2012

2.5 Modeling Approach

The random forest (RF) algorithm26,27 is an ensemble classifier that fits decision trees (models) tosubsamples of training data, then combines predictions from the resulting trees. The algorithmbootstrap samples the training data and approximately 63% of the original observations are usedat least once in model development. Data not included in the bootstrapped training sample arecalled out-of-bag observations. In addition to the randomness of the bootstrapping, the numberof variables sampled at each decision tree node is randomly selected from a subset of predictors;by default the number tried (mtry) is the square root of all available predictors.27 In this study,mtry is parameterized using the tuning function provided in the R implementation of RF by Liawet al.38 and Prasad et al.39 as is commonplace for RF applications with many predictor variables.The variable importance score is used to assess the contribution made by each of the spectral,

Table 2 Vegetation indices used in this study.

Vegetation indices

NDVI ¼ ðNIR−REDÞðNIRþREDÞ

mNDVI ¼ ðNIR−REDÞðNIRþREDÞ−ð2×BLUEÞ

REEI ¼ REDNIR

mEVI ¼ ðNIR−REDÞðNIRþC1×REDÞ−ðC2×BLUEþLÞ ð1þ LÞ

where: G ¼ 2.5, C1 ¼ 6.0, C2 ¼ 7.5, L ¼ 10

SAVI ¼ ð1þLÞ×ðNIR−REDÞðNIRþREDþLÞ

where: L ¼ 0.5

mARVI ¼ NIR−RBNIRþRB

where: RB ¼ RED − γðBLUE − REDÞ; γ ¼ 2.0; L ¼ 0.5

RGRI ¼ REDGREEN

mARI1 ¼ ðGREEN−1ÞðRED−1Þ

mARI2 ¼ NIR ðGREEN−1ÞðRED−1Þ

TVI ¼ 0.5ð120ðNIR −GREENÞ − 200ðRED −GREENÞ

Note: NDVI = normalized difference vegetation index, mNDVI = modified normalized difference vegetationindex, REEI = red edge extrema index, mEVI = modified enhanced vegetation index, SAVI = soil adjustedvegetation index, mARVI = modified atmospherically resistance vegetation index, RGRI = red green ratioindex, mARI1 = anthocyanin reflectance index 1, mARI2 = anthocyanin reflectance index 1, and TVI = trans-formed vegetation index.

Table 3 Textural indices used in this study.

Name Equation

Homogeneity HOM ¼ PN−1i ;j¼0

Pi;j

1þði−jÞ2

Contrast CON ¼ PN−1i ;j¼0 Pi;j ði − jÞ2

Entropy ENT ¼ PN−1i ;j¼0 Pi;j ð− ln Pi−j Þ

Difference Entropy DEN ¼ PN−1k¼0 Pk ð− ln Pk Þ

Variance VAR ¼ σ2i ;jPN−1

i ;j¼0 Pi;j ði ; j − μi ;j Þ where: μi−j ¼PN−1

i ;j¼0Pi;j

N2

Correlation COR ¼ PN−1i ;j¼0½ði − μi Þðj − μj Þ

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðσ2i Þðσ2j Þ

q�

Note: Pi;j represents the intensities of the pixel samples of i and j in the moving window (matrix), their spatiallocation depends on the cardinal direction of the gray level covariance matrix parameterization.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-5 Vol. 6, 2012

texture, and geometric remotely sensed variables made in the resultant RF classification models.Variable importance in RF is the normalized difference classification accuracy of each predictor,both when it is included as an observation and when it has been randomly permuted.26,27

RF was chosen because it has been shown to have superior predictive power in a number ofclassification studies39–43 and validated against a number of statistical modeling approaches forecological43 and vegetation health modeling applications.8

The 2008 vegetation, texture, and geometric indices were compared to a range of stratifica-tions of the combined TCHI (data not shown). TCHI stratifications of less than 10 point precision(i.e., 10/100) were not considered robust given the human error inherent in the underling crowncondition assessments. Two classifications of the in situ crown condition indices were used in thefinal analysis. The first reduced the THCI score to 10% increments, resulting in a 10 classvariable. This level of aggregation is consistent with the estimated error of the measurementprotocols discussed earlier. We recognize that as a practical management tool most forestcondition assessment indices are applied across a five class scale, so a second analysis wasundertaken which reduced the THCI score to 20% increments to give five classes.

To understand the robustness of the developed models, the relative size of the model devel-opment versus model testing datasets was varied from 10% to 100% with 10% indicative of amodel which was developed using 10% of the data and tested on the remaining 90%, and a 100%model indicative of a model developed using all observations. Each RF was run 100 timesfor each split and the out-of-bag error (OOBE), and weighted Cohen’s kappa coefficients(κ) were calculated (all were significant p < 0.01), comparing the models predictions againstthe observations. κ is a statistical measure of inter-rater agreement between categorical dataand accounts for a hypothetical probability of chance.44,45 Accordingly, κ ¼ þ1 represents

Fig. 2 Map of site 8 featuring (a) the delineated tree crowns, and (b) and (c) two examples oftextural crown matrices from which mean crown textural indices were derived. The exampleindices shown for crown 801 are (b) correlation (COR) for the 780 nm (NIR) band and (c) contrast(CON) for the 450 nm (BLUE) band and each group contains an image from 2007, 2008, and 2010.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-6 Vol. 6, 2012

complete agreement between observations and predictions and κ ¼ −1 suggests there is no moreagreement other than one would expect by chance. The Landis and Koch46 κ threshold scale(Table 4) provides a benchmark for interpreting the magnitude of κ in the context of the pre-dictive strength.

Once the RF model was developed and assessed, the 100% data model was applied to the2007 and 2010 imagery and changes in the crown condition index assessed.

3 Results

The κ scores were captured for the two models (Fig. 3). A score is given for each of the 10training and validation fractions (10% to 100% in 10% increments) of the available data(n ¼ 80 tuart crowns across the 20 sites). As expected, κ increases as the fraction of dataused for model development grows larger. Figure 4, a companion to Fig. 3, shows the corre-sponding OOBE scores for the two models and expectedly, OOBE becomes smaller as thefraction of data used increases. In order to determine an ideal mtry parameterization, thetrainRF() method38 was used, and this resulted in mtry ¼ 6 for all models except the 10class 30% and the 20 class 90% and 100% groups which were mtry ¼ 3.

Analysis of variable importance extracted from the 2008 100% training models (Fig. 5)shows that the spectral and textural indices containing the 780 nm spectral band were mostoften selected in model development. The geometric index of the size of the crown was nota top 5 variable selected by any model. Generally, spectral indices containing the 675 and780 nm bands were dominant. Although less frequently selected, the 450 and 550 nm spectralbands were also important in both models.

Figures 6 and 7 show the distribution of remote-sensing-derived THCI values of the 80crowns for 2007, 2008, and 2010 using all the available 2008 data for model development.A shift of the distribution to lower classes (smaller number) is indicative of a decline incrown condition, whereas conversely a shift of the distribution to healthier classes (larger num-ber) indication of a recovery of condition. For example, Fig. 6 shows some recovery between2007 and 2008, with 10% of the crowns populating the healthiest class in 2008 that were notclassified as class 5 in 2007.

Between 2008 and 2010 crowns contracted from both the least and most healthy classes topopulate moderate health classes. The net result could be considered a recovery with the majorityof crowns healthy, 60% were classified as class 4 (5 class model refer Fig. 6) and class 7 (10 classmodel refer Fig. 7) which represents a change of almost 20% from 2008 to 2010. Indeed, theleast healthy classes 1 (Fig. 6) and 1 and 2 (Fig. 7) show continuous recovery between 2007 to2008 but not complete recovery over this period. This suggests a recovery from disturbances(prior to 2007) can take a number of years to manifest. Conversely, the increased frequency ofmoderately healthy classes contributed by both additional healthy and declining crowns isindicative of the resilience of the species to dealing with, and recovering from dieback overmany years.

Table 4 Weighted Cohen’s kappa coefficient strength of agreement bench-marks after Landis and Koch.46

Weighted Cohen’s kappa Agreement

−1.00 to 0.00 Poor

0.00 to 0.20 Slight

0.21 to 0.40 Fair

0.41 to 0.60 Moderate

0.61 to 0.80 Substantial

0.81 to 1.00 Almost Perfect

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-7 Vol. 6, 2012

4 Discussion

The remote sensing driven eucalypt crown condition indicator presented here provides insightinto the extent and severity of tuart decline in Yalgorup National Park between 2007 and 2010.Crown condition shifts from the less healthy classes to healthier classes (2007 to 2008) andretracts back from the healthiest classes to moderately healthy classes (2008 to 2010). In thelast year of observation, 2010, the majority of tuart (60%) are predicted to be of moderate health.These results are indicative of a combination of mild stressors impacting the condition of thetuart, a finding supported by recent evidence of pathogens6 and soil bacterial communities,9 bothslow-acting agents, being linked to these declines. Eucalypt crown decline often occurs over anumber of years rather than instantaneously. However, severe weather, in particular droughtstress, pests, and some diseases can cause rapid defoliation events from time to time. Generally,the condition of the dominant species is a surrogate for ecosystem condition health24 with crowncondition becoming widely used as an indicator in land management policy for describingaspects of aesthetics, production, and the biodiversity of ecosystems.47 Remote sensing driventechniques are increasingly useful in providing data to support the development of these criticalindicators.1,8

These results confirm an enhancement of the eucalypt crown condition indicator from Evanset al.25 and the RF variables importance results (Fig. 6) indicate how this was done. Spectral andtextural indices containing the 780 nm spectral band were most often selected, an expected resultgiven the evidence that near infrared reflection is sensitive to changes in foliage condition.48–50

All of the VIs most commonly selected by the models utilized the 780 nm band in addition to675 nm (red) which is a strong indicator of chlorophyll absorption. VI containing 450 nm (blue)and 550 nm (green) were also important, and this suggests that variations in this part of theelectromagnetic spectrum, for example green reflectance and blue absorption, are being driven

10 20 30 40 50 60 70 80 90 100

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Percent of 80 crowns used for model development

Wei

gh

ted

Co

hen

’s k

app

a

10 Class model5 Class model

Fig. 3 Comparison of the weighted Cohen’s kappa for the 10 and 5 class models for each of thetraining and validation fractions used in the model development.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-8 Vol. 6, 2012

by changes in foliage condition. This strong relationship between the blue and green spectralbands is known to explain variance in crown chlorophyll content51–56 and as a result of a shift inthe location and position of the red edge from the red-nir to blue-green part of the electromag-netic spectrum as chlorophyll content decreases from leaf senescence or absence.24,57,58

An interesting finding of this study is that as the number of increment classes changed fromsmall (5 class) to large (10 class) the tendency of the RF to select texture-based indices increased.This result implies that texture-based measures allows for finer discrimination of more levels ofcrown condition whereas spectral-based indices allow for more broad scale canopy assessments.

The RF modeling approach used in this study allowed us to investigate two key aspects of thecrown condition dataset. First, as the RF modeling procedure is not computationally expensive,iterative modeling and the bootstrapping approach provides insight into the robustness of thedeveloped models. The results indicate, as expected, that as a higher proportion of data is used inmodel development the overall accuracy of the model improves in both the 5 and 10 class data-sets. Second, a typical jackknifing of the dataset into 70% for model development and 30% forindependent model prediction in this study produces κ of 0.67 for the 10 class and 0.64 for the 5class models, a substantial agreement on the Landis and Koch46 benchmark (Table 4), andsimilarly promising to reported accuracies of Evans et al.25 who applied standard linear regres-sion of mean crown VI to classify condition. Both models achieve a κ of 0.84 when using 90% ofthe observations and κ ¼ 1 when all of the available data are used. It is important to note thatgiven RF internally bootstraps, this means that when the models are developed using only 90%of the observations RF is actually using as little as 45 crowns.

Whilst RF is not a descriptive or inferential tool it is useful for identifying important ecologicalvariables for interpretation.43,59 We have used RF variable importance to isolate useful spectral andtextural indicators of crown health through classification and prediction of a categorical set of

10 20 30 40 50 60 70 80 90 100

020

4060

8010

0

Percent of 80 crowns used for model development

Ou

t−o

f−b

ag e

rro

r (%

)

10 Class model5 Class model

Fig. 4 Comparison of the random forest out-of-bag error score for the 10 and 5 class models foreach of the training and validation fractions used in the model development.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-9 Vol. 6, 2012

crown condition classes of the TCHI. RF is non-parametric, requires no distributional assumptions,and can handle the ‘small n large p’ scenario demonstrated in this study, by the splitting theavailable data into incremental portions.27,38 RF offers a robust, efficient, and accurate alternativeto parametric and semi-parametric methods for identifying ecological indicators.

As expected, when comparing the 10 to the 5 class datasets the results indicate the OOBE, astandard error indicator of RF’s internal predictive capacity, is higher for the 10 class whencompared to the 5 class model confirming that a smaller number of crown condition classesproduces a more accurate overall model.27,38,39,60–63 The trade-off between classificationprecision (number of classes) and model accuracy (OOBE and κ) is critical to its applicationas an indicator where it is likely a lower precision classification is sufficient for forest healthmanagement, given limitations to response options, and therefore greater accuracy wouldcontribute to greater confidence in its application.

The development of remotely derived eucalypt crown condition indicators is imperative forthe management of the large forested estates and to quantify changes in structure and health overtime. High spatial and spectral remotely sensed data is a rich source of information for forestmanagers and as demonstrated in these results, can make accurate predictions of crown conditionacross a large number of plots and crowns. Critical to the success of this type of remote sensingdata is the capacity to explicitly model changes in the biophysical properties of tree crowns, suchas a reduction in chlorophyll and changes in crown structure such as defoliation.

In terms of the application of this technique to forest and woodland management the 5 classassessment could offer broad insights to overall canopy condition whereas the 10 class variable

Variable Importance0.8 0.9 1.0 1.1 1.2

SAVI

NDVI

Mean COR 780nm

MARI2

REEI

mAVRI

REEI

NDVI

mNDVI

Mean HOM 550nm

10 Class model

5 Class model

Fig. 5 Top 5 variable importance from the models used to estimate the 2007 and 2010TCHI scores using all available in situ data (n ¼ 80) from 2008. SAVI ¼ soiladjusted vegetation index, NDVI ¼ normalized difference vegetation index, Mean COR 780 nmis the mean correlation of the 780 nm (nir) spectral band, MARI2¼anthocyaninreflectance index2, REEI¼ rededgeextrema index, mAVRI ¼ modified atmospherically resistantvegetation index, mNDVI ¼ normalized difference vegetation index, and mean HOM 550 nm isthe mean homogeneity of the 550 nm (green) spectral band.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-10 Vol. 6, 2012

could offer potentially increased sensitivity to internal crown structure and condition.This additional sensitivity would be valuable, for example, when looking for evidence of theinitial stages of eucalypt crown dieback and discerning the different agents involved in thedeclines.

As discussed, in situ crown condition data collection is an expensive task and as a resultremote sensing and modeling approaches could represent a potential cost saving throughboth limiting the number of site visits and reducing the sample size of crowns assessed inthe field. We recognize that the assessment of 80 crowns across the national park for modeldevelopment is relatively small and that, as this work continues, the collection of additional

(a) 2007

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

1 2 3 4 5

(b) 2008

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

1 2 3 4 5

(c) 2010

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

1 2 3 4 5

Fig. 6 Histograms of the total crown health index (TCHI) classification using the 5 class modelwhere 1 ¼ least healthy or declining (TCHI of 1% to 20%) and 5 ¼ most healthy (TCHI of 81% to100%) predictions for (a) 2007 and (c) 2010; and (b) the in situ observations for 2008. All percen-tages calculated using all available data (n ¼ 80).

(a) 2007

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

2 4 6 8 10

(b) 2008

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

2 4 6 8 10

(c) 2010

Class

Per

cent

age

of to

tal

10

20

30

40

50

60

2 4 6 8 10

Fig. 7 Histograms of the total crown health index (TCHI) classification using the 10 class modelwhere 1 ¼ least healthy or declining (TCHI of 1% to 10%) and 10 ¼ most healthy (TCHI of 91% to100%) predictions for (a) 2007 and (c) 2010; and (b) the in situ observations for 2008. All percen-tages calculated using all available data (n ¼ 80).

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-11 Vol. 6, 2012

in situ data is imperative, especially to test the predictions through time. Evans et al.25 and Hortonet al.,30 however, indicated that information on 80 tuart crowns was a suitable number for theassessment of canopy condition from the ground. In addition the capacity of the RF approach toallow investigation of both the number of classes, and the robustness of the relationshipsprovides confidence in these results.

Finally, the collection of additional in situ data and the use of remotely sensed mean crownstatistics in the RF models required labor-intensive and expensive manual delineation of crownsboth in the field and within the imagery. Future work will use object-based image analysis (OBIA)software that has been demonstrated as an alternative to delineating crowns manually using pro-cesses such as feature extraction or segmentation. This process has been demonstrated on non-eucalypt biomes64–66 and eucalypt biomes,67 but it has not been conducted at the crown scale onnative eucalypts exhibiting advanced crown decline and will therefore present some challenges.

5 Conclusion

This study has demonstrated the application of RF models trained using remotely sensed spec-tral, textural, and geometric indices to predict eucalypt crown condition and its change throughtime. The results indicate that two models, using 5 and 10 descriptor classes were accurate androbust and could represent a cost-effective way of monitoring and quantifying crown conditionof tuart over large areas within an operational environment. The results confirm that spectral andtextural information in the 450, 550, 675, and 780 nm spectral bands was critical in the modelpredictions and texture indices were more informative when increasing the sensitivity of themodels to smaller variations in crown condition. Once validated, the models confirmed the over-all condition of tree crowns within Yalgorup National Park declines from 2007 to 2010. Thisstudy has shown the potential that exists for the use of this indicator of eucalypt crown conditionindicator for monitoring changes in the crown condition of tuart and other eucalypt species innative and plantation forests.

Acknowledgments

We are grateful for the funding from the Australian Research Council (LP0668195) Departmentof Conservation and Environment, Alcoa of Australia, Mandurah City Council, and BemaxCable Sands. We thank SpecTerra Services Ltd. for their substantial support in the provisionof the orthorectified digital multispectral imagery utilized for this study. We also thank Dr.Frank Honey and Professor Nicholas Coops and the anonymous reviewers for their feedbackat various stages. The work was supported by iVEC (http://www.ivec.org) through the use ofadvanced computing resources provided by the Australian Resources Research Centre Facilitylocated at Technology Park, Perth. The work of Bradley Evans is supported by a Murdoch Uni-versity PhD scholarship and a Centre of Excellence for Climate Change Woodland and ForestHealth scholarship.

References

1. N. Coops et al., “The development of a Canadian dynamic habitat index using multi-tem-poral satellite estimates of canopy light absorbance,” Ecol. Indic. 8(5), 754–766 (2008),http://dx.doi.org/10.1016/j.ecolind.2008.01.007.

2. G. N. Bastin et al., “Indicators of landscape function?: comparing patchiness metrics usingremotely-sensed data from rangelands,” Ecol. Indic. 1(4), 247–260 (2002), http://dx.doi.org/10.1016/S1470-160X(02)00009-2.

3. V. Jurskis, “Eucalypt decline in Australia, and a general concept of tree decline and dieback,”Forest Ecol. Manag. 215(1–3), 1–20 (2005), http://dx.doi.org/10.1016/j.foreco.2005.04.026.

4. R. J. Harper, K. R. J. Smettem, and R. J. Tomlinson, “Using soil and climatic data to esti-mate the performance of trees, carbon sequestration and recharge potential at the catchmentscale,” Aust. J. Exp. Agr. 45(11), 1389 (2005), http://dx.doi.org/10.1071/EA04186.

2

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-12 Vol. 6, 2012

5. R. J. Hooper and K. Sivasithamparam, “Characterization of damage and biotic factors asso-ciated with the decline of Eucalyptus wandoo in southwest Western Australia,” Can. J.Forest Res. 35(11), 2589–2602 (2005), http://dx.doi.org/10.1139/x05-162.

6. P. M. Scott et al., “Patheogenicuty of Phytopthora multivora to Eucalyptus gomphocephalaand E. marginata,” For. Pathol. 42(4), 289–298 (2012), http://dx.doi.org/10.1111/j.1439-0329.2011.00753.x.

7. D. C. Close et al., “Does low-intensity surface fire increase water- and nutrient- availabilityto overstorey Eucalyptus gomphocephala?,” Plant Soil 349(1-2), 203–214 (2011), http://dx.doi.org/10.1007/s11104-011-0862-3.

8. A. Haywood and C. Stone, “Mapping eucalypt forest susceptible to dieback associated withbell miners (Manorina melanophys) using laser scanning, SPOT 5 and ancillary topogra-phical data,” Ecol. Model. 222(5), 1174–1184 (2011), http://dx.doi.org/10.1016/j.ecolmodel.2010.12.012.

9. Y. F. Cai et al., “Soil bacterial functional diversity is associated with the decline of Euca-lyptus gomphocephala,” Forest Ecol. Manag. 260(6), 1047–1057 (2010), http://dx.doi.org/10.1016/j.foreco.2010.06.029.

10. R. D. Archibald, “Fire and the persistence of tuart woodlands,” Murdoch University, Perth(2006).

11. R. D. Archibald et al., “Understorey thinning and burning trials are needed in conservationreserves: the case of tuart (Eucalyptus gomphocephala DC),” Ecol. Manag. Restor. 11(2),108–112 (2010), http://dx.doi.org/10.1111/(ISSN)1442-8903.

12. D. C. Close, “Premature decline of eucalyptus and altered ecosystem processes in theabsence of fire in some Australian forests,” The Bot. Rev. 75(2), 191–202 (2009), http://dx.doi.org/10.1007/s12229-009-9027-y.

13. C. Stone et al., “Harmonisation of methods for the assessment and reporting of forest healthin Australia—a starting point A discussion paper prepared by a sub-committee of the ForestResearch Working Group on Forest Health,” Austral. For. 66(4), 233–246 (2003).

14. N. J. Souter et al., “Evaluation of a visual assessment method for tree condition of eucalyptfloodplain forests,” Ecol. Manag. Restor. 11(3), 210–214 (2010), http://dx.doi.org/10.1111/emr.2010.11.issue-3.

15. J. T. Kerr and M. Ostrovsky, “From space to species: ecological applications for remotesensing,” Trends Ecol. Evol. 18(6), 299–305 (2003), http://dx.doi.org/10.1016/S0169-5347(03)00071-5.

16. N. C. Coops et al., “Assessment of crown condition in eucalypt vegetation by remotelysensed optical indices,” J. Environ. Qual. 33(3), 956–964 (2004), http://dx.doi.org/10.2134/jeq2004.0956.

17. S. J. Zarnoch, W. A. Bechtold, and K. W. Stolte, “Using crown condition variables as indi-cators of forest health,” Can. J. Forest Res. 34(5), 1057–1070 (2004), http://dx.doi.org/10.1139/x03-277.

18. P. Caccetta et al., “Continental monitoring: 34 years of land cover change using Landsatimagery,” in 32nd Int. Symp. on Remote Sensing of Environment, pp. 25–29 (2007).

19. J. F. Wallace, P. A. Caccetta, and H. T. Kiiveri, “Recent developments in analysis of spatialand temporal data for landscape qualities and monitoring,” Austral Ecol. 29(1), 100–107(2004), http://dx.doi.org/10.1111/j.1442-9993.2004.01356.x.

20. A. Pinar and P. J. Curran, “Technical note grass chlorophyll and the reflectance red edge,” Int.J. Remote Sens. 17(2), 351–357 (1996), http://dx.doi.org/10.1080/01431169608949010.

21. K. M. Barry, C. Stone, and C. L. Mohammed, “Crown-scale evaluation of spectral indicesfor defoliated and discoloured eucalypts,” Int. J. Remote Sens. 29(1), 47–69 (2008), http://dx.doi.org/10.1080/01431160701281056.

22. E. Pietrzykowski et al., “Effects of Mycosphaerella leaf disease on the spectral reflectanceproperties of juvenile Eucalyptus globulus foliage,” Forest Pathol. 36(5), 334–348 (2006),http://dx.doi.org/10.1111/efp.2006.36.issue-5.

23. C. E. Woodcock and A. H. Strahler, “The factor of scale in remote sensing,” Remote Sens.Environ. 21(3), 311–332 (1987), http://dx.doi.org/10.1016/0034-4257(87)90015-0.

24. C. Stone and A. Haywood, “Assessing canopy health of native eucalypt forests,” Ecol.Manag. Rest. 7(June), 24–30 (2006), http://dx.doi.org/10.1111/j1442-8903.2006.00288.x.

3

4

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-13 Vol. 6, 2012

25. B. J. Evans et al., “Dieback classification modelling using high resolution digital multi spec-tral imagery and in situ assessments of crown condition,” Remote Sens. Lett. 3(6), 541–550(2012), http://dx.doi.org/10.1080/01431161.2011.639400.

26. L. Breiman, “Bagging predictors,” Mach. Learn. 24(2), 123–140 (1996).27. L. Breiman, “Random forests,”Mach. Learn. 45(1), 5–32 (2001), http://dx.doi.org/10.1023/

A:1010933404324.28. E. M. Heddle, O. W. Loneragan, and J. J. Havel, “Vegetation complexes of the darling

system,” in Atlas of Natural Resources: Darling System, pp. 33–74, Perth (1980).29. C. Portlock et al., Yalgorup National Park management plan 1995–2005, Perth, pp. 101–

102 (1993).30. B. M. Horton et al., “Crown condition assessment: an accurate, precise and efficient method

with broad applicability to Eucalyptus,” Austral. Ecol. 36(6), 709–721 (2011), http://dx.doi.org/10.1111/j.1442-9993.2010.02206.x.

31. USDAFS, “Forest inventory and analysis: national core field guide, 1: field data collectionprocedures for phase 2 plots, version 1.6.,” US Department of Agriculture Forest Service,Forest Inventory and Analysis, Washington D. C. (2002).

32. M. E. Schomaker et al., “Crown-condition classification?: a guide to data collection andanalysis,” Gen. Tech. Rep. SRS-102, Asheville, NC, p. 78 (2007), http://www.treesearch.fs.fed.us/pubs/27730.

33. G. A. Kile, C. R. A. Turnbull, and F. D. Podger, “Effect of ‘regrowth dieback’ on someproperties of Eucalyptus obliqua trees,” Aust. Forest Res. 11, 55–62 (1981).

34. T. J. Wardlaw, “Management of Tasmanian forests affected by regrowth dieback,” New Zeal.J. For. Sci. 19, 265–276 (1989).

35. S. Furby and N. Campbell, “Calibrating images from different dates to ‘like-value’ digitalcounts,” Remote Sens. Environ. 77(2), 186–196 (2001), http://dx.doi.org/10.1016/S0034-4257(01)00205-X.

36. R. M. Haralick, K. Shanmugam, and I. H. Dinstein, “Textural features for image classifica-tion,” IEEE Trans. Syst., Man, Cybern. 3(6), 610–621 (1973), http://dx.doi.org/10.1109/TSMC.1973.4309314.

37. R. M. Haralick, “Statistical and structural approaches to texture,” Proc. IEEE 67(5), 786–804 (1979), http://dx.doi.org/10.1109/PROC.1979.11328.

38. A. Liaw and M. Wiener, “Classification and regression by randomforest,” R News 2(3),1822 (2002).

39. A. M. Prasad, L. R. Iverson, and A. Liaw, “Newer classification and regression tree tech-niques: bagging and random forests for ecological prediction,” Ecosystems 9(2), 181–199(2006), http://dx.doi.org/10.1007/s10021-005-0054-1.

40. A. Bosch, A. Zisserman, and X. Muoz, “Image classification using random forests andferns,” in 11th Int. Conf. Computer Vision IEEE , pp. 1–8 (2007).

41. R. Lawrence, S. Wood, and R. Sheley, “Mapping invasive plants using hyperspectral ima-gery and Breiman Cutler classifications (randomforest),” Remote Sens. Environ. 100(3),356–362 (2006), http://dx.doi.org/10.1016/j.rse.2005.10.014.

42. M. Pal and P. M. Mather, “An assessment of the effectiveness of decision tree methods forland cover classification,” Remote Sens. Environ. 86(4), 554–565 (2003), http://dx.doi.org/10.1016/S0034-4257(03)00132-9.

43. D. R. Cutler et al., “Random forests for classification in ecology,” Ecol. 88(11), 2783–2792(2007), http://dx.doi.org/10.1890/07-0539.1.

44. J. Cohen, “A coefficient of agreement for nominal scales,” Educ. Psychol. Meas. 20(1), 37–46 (1960) http://dx.doi.org/10.1177/001316446002000104.

45. J. Cohen, “Weighted kappa: nominal scale agreement with provision for scaled disagree-ment or partial credit,” Psychol. Bull. 70(4), 213–220 (1968), http://dx.doi.org/10.1037/h0026256.

46. J. R. Landis and G. G. Koch, “An application of hierarchical kappa-type statistics in theassessment of majority agreement among multiple observers,” Biometrics 33(2), 363–374(1977), http://dx.doi.org/10.2307/2529786.

47. D. Keith and E. Gorrod, “The meanings of vegetation condition,” Ecol. Manag. Rest. 7(s1),S7–S9 (2006), http://dx.doi.org/10.1111/j.1442-8903.2006.00285.x.

5

6

7

8

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-14 Vol. 6, 2012

48. N. C. Coops et al., “Chlorophyll content in eucalypt vegetation at the leaf and canopy scalesas derived from high resolution spectral data,” Tree Physiol. 23(1), 23–31 (2003), http://dx.doi.org/10.1093/treephys/23.1.23.

49. N. Coops et al., “Comparison of green leaf eucalypt spectra using spectral decomposition,”Aust. J. Bot. 50(5), 567 (2002) http://dx.doi.org/10.1071/BT01082.

50. C. Stone, L. Chisholm, and N. C. Coops, “Spectral reflectance characteristics of eucalyptfoliage damaged by insects,” Aust. J. Bot. 49, 687–698 (2001), http://dx.doi.org/10.1071/BT00091.

51. G. A. Carter and A. K. Knapp, “Leaf optical properties in higher plants: linking spectralcharacteristics to stress and chlorophyll concentration,” Am. J. Bot. 88(4), 677–684 (2001),http://dx.doi.org/10.2307/2657068.

52. G. S. Rabideau, C. S. French, and A. S. Holt, “The absorption and reflection spectra ofleaves, chloroplast suspensions, and chloroplast fragments as measured in an Ulbrichtsphere,” Am. J. Bot. 33(10), 769–777 (1946), http://dx.doi.org/10.2307/2437271.

53. R. A. Moss and W. E. Loomis, “Absorption spectra of leaves I. the visible spectrum,” PlantPhysiol. 27(2), 370–391 (1952), http://dx.doi.org/10.1104/pp.27.2.370.

54. H. K. Lichtenthaler, Applications of Chlorophyll Flourescence, p. 149, Klewer AcademicPublishers, Dordrecht, The Netherlands (1987).

55. H. K. Lichtenthaler, “Vegetation stress: an introduction to the stress concept in plants,” J.Plant Physiol. 148(1), 4–14 (1996), http://dx.doi.org/10.1016/S0176-1617(96)80287-2.

56. H. K. Lichtenthaler, “Chlorophylls and carotenoids?: Measurement and characterization byUV-VIS,” Anal. Chem. 1–8 (2001), http://dx.doi.org/10.1002/0471142913.faf0403s01.

57. A. Pinar and P. J. Curran, “Technical Note Grass chlorophyll and the reflectance red edge,”Int. J. Remote Sens. 17(2), 351–357 (1996), http://dx.doi.org/10.1080/01431169608949010.

58. R. A. Jago, M. E. J. Cutler, and P. J. Curran, “Estimating canopy chlorophyll concentrationfrom field and airborne spectra,” Remote Sens. Environ. 68(3), 217–224 (1999), http://dx.doi.org/10.1016/S0034-4257(98)00113-8.

59. C. Strobl et al., “Conditional variable importance for random forests,” BMC Bioinf. 9, 307(2008), http://dx.doi.org/10.1186/1471-2105-9-307.

60. V. Svetnik et al., “Random forest: a classification and regression tool for compound clas-sification and QSAR modelling,” J. Chem. Inf. Comput. Sci. 43(6), 1947–1958 (2003),http://dx.doi.org/10.1021/ci034160g.

61. R. Díaz-Uriarte and S. A. de Andrés, “Gene selection and classification of microarray datausing random forest,” BMC Bioinf. 7, 3 (2006), http://dx.doi.org/10.1186/1471-2105-7-3.

62. P. Gislason, J. Benediktsson, and J. Sveinsson, “Random forests for land cover classifica-tion,” Pattern Recogn. Lett. 27(4), 294–300 (2006) http://dx.doi.org/10.1016/j.patrec.2005.08.011.

63. J. S. Evans and S. A. Cushman, “Gradient modelling of conifer species using random forests,”Landscape Ecol. 24(5), 673–683 (2009), http://dx.doi.org/10.1007/s10980-009-9341-0.

64. K. Johansen et al., “Mapping riparian condition indicators in a sub-tropical savanna envir-onment from discrete return LiDAR data using object-based image analysis,” Ecol. Indic.10, 796–807 (2010), http://dx.doi.org/10.1016/j.ecolind.2010.01.001.

65. Q. Yu et al., “Object-based detailed vegetation classification with airborne high spatial reso-lution remote sensing imagery,” Photogramm. Eng. Remote Sens. 72(7), 799–811 (2006).

66. Y. Ke, L. J. Quackenbush, and J. Im, “Synergistic use of QuickBird multispectral imageryand LIDAR data for object-based forest species classification,” Remote Sens. Environ.114(6), 1141–1154 (2010), http://dx.doi.org/10.1016/j.rse.2010.01.002.

67. A. Haywood and C. Stone, “Semi-automating the stand delineation process in mappingnatural eucalypt forests,” Austral. For. 74(1), 13–22 (2011).

Biographies and photographs of the authors are not available.

9

10

11

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-15 Vol. 6, 2012

Queries

1. Corresponding address taken from database. Is it okay?

2. This query was generated by an automatic reference checking system. The references 13, 26,33, 34, 38, 65, 67 could not be located in the databases used by the system. While thesereferences may be correct, we ask that you check them so we can provide as many links tothe referenced articles as possible.

3. Please provide information that tells what kind of reference Ref. 10 is.

4. Please provide publisher name and location for Ref. 18.

5. Please provide publisher name for Ref. 28 and 29.

6. Please provide issue number for Refs. 33, 34, 50, 59, 61, 64.

7. Please provide page range for Ref. 38.

8. Please provide publisher name and location for Ref. 40.

9. A check of online databases revealed a possible error in this reference. (53) The year hasbeen changed from '1951' to '1952'. Please confirm this is correct.

10. Please provide volume and issue number for Ref. 56.

11. This article currently does not include biographies or photos for any of the authors. If youwish to include biographies, please provide text and photos with your corrections.

Evans et al.: Enhancing a eucalypt crown condition indicator driven by high spatial . . .

Journal of Applied Remote Sensing XXXXXX-16 Vol. 6, 2012