FH

Ra

b

a

ARA

KHHSHSS

1

1

si2iegbsTaviH(

(

0h

International Journal of Applied Earth Observation and Geoinformation 28 (2014) 140–149

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

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

orest tree species discrimination in western Himalaya using EO-1yperion

ajee Georgea, Hitendra Padaliab,∗, S.P.S. Kushwahab

Department of Environment and Forests, Andaman and Nicobar Islands, Port Blair 744102, IndiaForestry and Ecology Department, Indian Institute of Remote Sensing, ISRO, Dehradun 248001, India

r t i c l e i n f o

rticle history:eceived 10 September 2013ccepted 21 November 2013

eywords:yperspectralyperionpecies discriminationimalayaVM

a b s t r a c t

The information acquired in the narrow bands of hyperspectral remote sensing data has potential to cap-ture plant species spectral variability, thereby improving forest tree species mapping. This study assessedthe utility of spaceborne EO-1 Hyperion data in discrimination and classification of broadleaved ever-green and conifer forest tree species in western Himalaya. The pre-processing of 242 bands of Hyperiondata resulted into 160 noise-free and vertical stripe corrected reflectance bands. Of these, 29 bands wereselected through step-wise exclusion of bands (Wilk’s Lambda). Spectral Angle Mapper (SAM) and Sup-port Vector Machine (SVM) algorithms were applied to the selected bands to assess their effectiveness inclassification. SVM was also applied to broadband data (Landsat TM) to compare the variation in classifi-cation accuracy. All commonly occurring six gregarious tree species, viz., white oak, brown oak, chir pine,

AM blue pine, cedar and fir in western Himalaya could be effectively discriminated. SVM produced a betterspecies classification (overall accuracy 82.27%, kappa statistic 0.79) than SAM (overall accuracy 74.68%,kappa statistic 0.70). It was noticed that classification accuracy achieved with Hyperion bands was signif-icantly higher than Landsat TM bands (overall accuracy 69.62%, kappa statistic 0.65). Study demonstratedthe potential utility of narrow spectral bands of Hyperion data in discriminating tree species in a hillyterrain.

. Introduction

.1. Species discrimination using remote sensing

Species-level information of the forests is essential for under-tanding their composition and distribution which in turn helpsn monitoring large areas for sustainable management (Sobhan,007). Conventionally, species discrimination for floristic mapping

nvolved exhaustive field work and is time-consuming as well asxpensive (Kent and Coker, 1992). Remote sensing is advanta-eously used in such a situation to discriminate different speciesased on their unique spectral properties that is attributed to thepecies biochemical and biophysical parameters (Clark et al., 2005).he variation in concentration of biochemical constituents in plantslso leads to variation in their spectral reflectance. However, subtleariations in reflectance of plants due to species taxonomic variabil-

ty are difficult to resolve using multispectral remote sensing data.yperspectral data sets due to their relatively narrow bandwidth

5–10 nm) and large number of spectral bands (about 100–250) as

∗ Corresponding author. Tel.: +91 0135 2524176.E-mail addresses: george rajee@yahoo.co.in (R. George), hitendra@iirs.gov.in

H. Padalia), spskushwaha@iirs.gov.in (S.P.S. Kushwaha).

303-2434/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.jag.2013.11.011

© 2013 Elsevier B.V. All rights reserved.

compared to larger bandwidth (70–400 nm) multispectral data sets(5–10 bands) (Schmidt and Skidmore, 2003), has ability to resolvesubtle absorption and reflectance features of plant species facilitat-ing their discrimination and mapping.

1.2. Dimensionality reduction of hyperspectral data

The higher spectral resolution increases the dimensionality ofhyperspectral data. Hence, it is absolutely necessary to reducethe number of bands and use only the optimal bands for furtherdata processing. Several dimensionality reduction techniques suchas principal component analysis (PCA), minimum noise fractiontransform (MNF), stepwise discriminant analysis (SDA), decisionboundary feature extraction (DBFE), sequential forward floating(SFF), steepest Ascent (SA), genetic algorithms (GA) (Asner andHeidebrecht, 2002; Landgrebe, 2003; Melgani and Bruzzone, 2004;Myint, 2001; Neville et al., 2003; Okin et al., 2001; Platt and Goetz,2004; Rashed et al., 2003) have been described. PCA reduces thedimensionality of data by segregating noise from set of uncorre-lated bands. MNF is a modified PCA which does data reduction

more effectively than PCA by producing principal components withdecreasing noise level. DBFE approach retains most of the origi-nal information while transforming the original feature space intoa space of lower dimensionality (Melgani and Bruzzone, 2004).

rth Ob

Slf(orTu(mRVifw(cas

1

pSdsApRmtSsb1sccdemdltwsdahat

1

spsctaepofi

ability and better solar illumination in the study area during thatperiod.

R. George et al. / International Journal of Applied Ea

FF minimize redundant contiguous bands applying local corre-ation criterion and subsequently uses a discriminative functionor transformation of data into lower dimensionality feature spaceGomez-Chova et al., 2003). SA algorithm is used for featureptimization from multivariate datasets. GA is iterative patternecognition technique for feature selection as well as classification.he optimal wavebands for target discrimination can be identifiedsing multivariate separability measures such as Wilk’s lambdaThenkabail, 2002). SDA is easy to use and is widely used opti-

al band selection method (Jain et al., 2007; Lucas et al., 2008;ay et al., 2010; Van Aardt and Wynne, 2001; Vyas et al., 2011).yas et al. (2011) identified 22 best Hyperion bands to discrim-

nate eight tropical vegetation classes of tropical dry deciduousorests in Gujarat, India. Thenkabail et al. (2004) applied step-ise discriminant analysis and recommended 22 best narrowbands

from 168 hyperspectral bands), in the 350–2500 nm range to dis-riminate natural vegetation species of shrubs, grasses, weeds, andgricultural crops based on the hyperspectral data from Africanavannah.

.3. Classification techniques for species mapping

Various classification methods, both parametric and non-arametric, are used for mapping species using hyperspectral data.pectral Angle Mapper (Vyas et al., 2011; Clark et al., 2005), Lineariscriminant analysis (Du and Ren, 2003; Clark et al., 2005); Deci-ion tree classifier (Lawrence et al., 2004; Pal and Mather, 2003),rtificial neural networks (Erbek et al., 2004; Foody, 2004), Sup-ort Vector Machine (Dalponte et al., 2009; Plaza et al., 2009) andandom forest (Chan and Palinckx, 2008) are some of the advancedethods of hyperspectral data classification. SAM and SVM are

he commonly used classifiers for hyperspectral data classification.AM determines the spectral similarity, compares unknown imagepectra with known endmember spectra by calculating the angleetween the corresponding vectors in feature space (Kruse et al.,993). SAM does not consider the absolute reflectance, but con-iders only the shape of different spectra (Waske et al., 2009), andontain the effect of topography and solar illumination on targetlassification. However, SAM is a non-iterative process, it thereforeoes not optimize classification accuracy from misclassified pix-ls (Soman et al., 2009). The iterative process based classificationethod such as SVM is therefore widely used for hyperspectral

ata classification. SVM is a supervised, non-parametric, machineearning technique. SVM does not depend on statistical data dis-ribution (Mountrakis et al., 2011). SVM approach performs wellith small training samples, even with data with hyper dimen-

ional feature space are classified, because it only considers trainingata close to the class boundary (Melgani and Bruzzone, 2004; Palnd Mather, 2006). Both SAM and SVM classification approachesave features applicable to target discrimination in hilly regionre therefore selected for comparison and development of efficientechnique.

.4. Tree species mapping in western Himalaya

In the temperate Himalaya, studies carried out using multi-pectral data have discriminated gregarious species, viz., chirine (Pinus roxburghii), cedar (Cedrus deodara) and broadleavedpecies viz., grey oak (Quercus leucotrichophora), brown oak (Quer-us semicarpifolia), etc., in their zone of dominance, primarilyaking advantage of their association with altitude and aspectlong with spectral information with reasonable accuracy. How-

ver, in the ecotonal zone (1750–2200 m amsl), it has not beenossible to discriminate intermingled patches of gregariouslyccurring species, viz., white oak, brown oak, rhododendrons,r, spruce (Picea smithiana), blue pine (Pinus wallichiana) solely

servation and Geoinformation 28 (2014) 140–149 141

based on spectral information without taking altitudinal refer-ence. Since vegetation-topographical associations are generallyinconsistent at local scale, classification incorporating topographicinformation results in arbitrary discrimination of intermixedspecies.

In this context, the present study was conducted to evalu-ate the potential of EO-1 Hyperion data for discrimination andclassification of forest tree species in western Himalaya. The sub-objectives of the study were: (i) to find out Hyperion wavebandsuseful for discrimination of gregarious forest tree species usingstep-wise discriminant analysis, (ii) to compare the performanceof SAM and SVM algorithms in classifying forest tree species,(3) to assess the improvement in classifying forest tree speciesusing narrow wavebands of Hyperion over broad bands of LandsatTM.

2. Study area

The study area, located between 30◦25′50′′ to 30◦30′36′′ N and77◦58′39′′ to 78◦04′51′′ E, (Fig. 1), stretches from Purola to Sankriin the higher reaches of western Himalaya in Uttarakhand Stateof India, covers an area of 250 km2. It includes Upper Yamunaand Tons Forest Divisions and also part of Govind Wildlife Sanctu-ary. The terrain is mountainous with altitude varying from 1600 mnear Purola to 3940 m msl near Sankri. Two major rivers, i.e. theYamuna and the Tons are noticeable in the area. The climate of thestudy area varies from sub-tropical at lower elevations to temper-ate and alpine at higher elevations. The rainfall varies from 1000 to1500 mm annually.

The study area was selected considering the presence of a rangeof tree species, characteristics of western Himalaya. Due to the vari-ation in rainfall, altitude and aspect, the study area supports threetype groups of forest viz., the sub-tropical pine forests, Himalayanmoist temperate and the Moist alpine scrub (Champion and Seth,1968). The major tree species include chir pine (P. roxburghii),grey oak (Q. leucotrichophora), brown oak (Q. semicarpifolia), cedar(C. deodara), fir (Abies pindrow), spruce (P. smithiana), rhododen-dron (Rhododendron arboretum), etc. The patches of temperateand alpine grasslands and rocky slopes also occur in the studyarea.

3. Data

The Earth Observation-1 (EO-1) Hyperion and Landsat TM dataof 2nd April 2010 and 3rd March 2010 respectively were down-loaded from the USGS website (URL: glovisusgs.com). Cloud-freesubsets of 33 km by 7.5 km from both the data were selected for thestudy. The EO-1 satellite is in an orbit that covers the same groundtrack as Landsat TM, approximately one minute later. This enablesto obtain images of the same ground areas at nearly same time,so that direct comparison of results can be made for Hyperion andLandsat TM. Hyperion sensor images the area at 30-m ground sam-ple distance over a 7.5 km swath and provides 242 spectral bandsfrom 357 to 2576 nm at 10 nm sampling interval. Hyperion imagewas procured in L1R (radiometrically corrected) and L1T (terraincorrected and georeferenced image) data format. Hyperion sceneof early summer (April month) was selected because of its highersignal to noise ratio owing to higher interspecies phenological vari-

The ancillary data viz., Survey of India topographic sheets (53J/1,53I/4, 53J/3, 53J/2) on 1:50,000 scale and latest forest working plansof Tons and Upper Yamuna Forest Divisions were also used forbackground information.

142 R. George et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 140–149

n of th

4

4

Tgrsvffaa

TT

Fig. 1. Locatio

. Methods

.1. Field data collection

The field surveys were carried out from January to March 2013.he spectral profiles of the gregarious tree species (Table 1), viz.,rey oak, brown oak, chir pine, blue pine, cedar, and fir wereecorded using SVC HR1024 hand-held spectroradiometer. Thepectra were collected with fibre optic light guide having field ofiew of 25◦ with the intention to cover the composite reflectance

rom tree canopy and background. The field spectra were collectedrom the trees taking advantage of topography and climbing ondjacent trees. Ten spectra for each species were recorded andveraged spectra were used for the analysis. Ground co-ordinates

able 1ree species attribute of the study area.

Tree species Characteristic feature

Grey oak (Quercus leucotrichophora) Large or medium sized tree, with leatherylong.

Brown oak (Quercus semicarpifolia) Attains height of 24–30 m. Leaves 5–12 ×young trees, often entire on older trees.

Chir pine (Pinus roxburghii) Large size tree reaching 30–50 m. Leaves

distinctly yellowish green.Blue pine (Pinus wallichiana) Large size tree between 30 m and 50 m in

five and are 12–18 cm long.Western Himalayan fir (Abies pindrow) Large tree growing to 40–60 m tall. Leave

the shoots.Cedar (Cedrus deodara) Large tree reaching 40–50 m tall. Leaves a

shoots, and in dense clusters of 20–30 on

e study area.

(latitude and longitude) of the sample locations were recordedusing Trimble Juno SB GPS receiver. The field spectra containing1024 bands at 1 nm interval were then resampled to the processed160 band Hyperion data using the wavelength and full width at halfmaximum (FWHM) information. In addition, a total 117 groundcontrol points with minimum 10 locations per species were col-lected following random sampling design to evaluate the speciesclassification accuracies.

4.2. Pre-processing of Hyperion and Landsat TM data

Many of the 242 bands of Hyperion are not usable because oflow signal to noise ratio. All the bands were individually visual-ized and those bands which were not calibrated or had very high

Phenology

dull green leaves, sharply toothed leaves and 6–16 cm Evergreen

2.5–7.5 cm, elliptical or oblong, spinous-toothed in Evergreen

are needle-like, very slender, 20–35 cm long, and Semi-deciduous

height. Leaves (“needles”) are in fascicles (bundles) of Evergreen

s are needle-like, 4–9 cm long and arranged spirally on Evergreen

re needle-like, 2.5–5 cm long, borne singly on long short shoots.

Evergreen

rth Ob

naEoafctwpLLtz

obvwctdfia

4

stai

A

wc

0imo

dttCf

stsdbttrb

4

dous

R. George et al. / International Journal of Applied Ea

oise were discarded. 160 bands were considered for discriminantnalysis. The atmospheric correction was carried out throughNVI’s (Environment for Visualizing Images ver. 4.5) Fast Line-f-Sight Atmospheric Analysis of Spectral Hyper-Cubes (FLAASH)lgorithm. It minimized the influence of the atmospheric effectsrom the image. Landsat TM was not subjected to atmosphericorrection but the northern part of the image having clouds andheir shadows was removed through masking. The L1R imageas preferred over the geo-corrected L1T image due to ease inre-processing, i.e. bad band removal and de-stripping, etc. Later,1R was georeferenced with respect to L1T image. Hyperion andandsat TM both were projected to Universal Transverse Merca-or (UTM) projection, World Geodetic System (WGS) 84 datum andone 44 N.

Hyperion image was subjected to pixel purity index (PPI) inrder to select locations in the image to derive multiple endmem-ers of each tree species class. Two spectral similarity measuresiz., Spectral Angle Mapper (SAM) and Spectral Feature Fitting (SFF)ere used to evaluate the effectiveness of the FLAASH atmospheric

orrection model for obtaining better results. The spectral analystool of ENVI was used to assess the similarities between imageerived species spectra and endmember averaged spectra from theeld. The endmember spectra of grassland were also included in thenalysis as it covers a major portion of the study area.

.3. Stepwise discriminant analysis

The dimensionality of Hyperion data was reduced through atep-wise discriminant analysis procedure based on Wilk’s lambdaest statistic. Wilk’s lambda is a multivariate analysis of variancend is found through a ratio of the determinants. Wilks’ lambda (A)s given by Green and Caroll (1978).

= |Seffect||Seffect| + |Serror| (1)

here S is a matrix which is also known as sum of squares (SS) andross-products (SSCP).

Wilks’ lambda values ranges from 0 to 1. The values close to indicating the group means are different and values close to 1

ndicating the group means are not different and 1 indicates alleans are same. The stepwise discriminant analysis was carried

ut using SPSS ver. 20.0 software.The multiple endmember average spectra of different species

erived from Hyperion reflectance image were subjected to con-inuum removal. After continuum removal spectra were subjectedo Wilk’s lambda discriminant analysis for optimal bands selection.ontinuum removal was carried out to choose absorption based

eatures from the spectral bands.In order to assess whether any information will be lost by

electing 29 optimum bands out of the processed 160 bands ofhe Hyperion data, both the image were classified by using theame endmembers using SVM algorithm. The spatial correspon-ence between two classified images was checked pixel by pixelasis using matrix function in Erdas Imagine Software. It was foundhat the classified map using 29 band image matched 95.55% withhe classified map using 160 band image. Since the classificationesults of both the images were almost similar, the image with 29ands was taken up for further classification.

.4. Evaluation of Hyperion spectra

The field spectra were compared to Hyperion endmember

erived spectra using the Spectral Angle Mapper algorithm and theutput is a ranked or weighted score for each pair of the spectrasing spectral analyst in ENVI. The above algorithm compares thepectra from field and image based on their spectral characteristics.

servation and Geoinformation 28 (2014) 140–149 143

The highest score indicates the closest match and indicates higherconfidence in the spectral similarity. SAM result less than or equalto minimum indicates a perfect match and receives a score of 1.SAM result greater than or equal to maximum receives a score of 0.

4.5. Spectral separability analysis

In order to ascertain that the multiple endmembers of differentspecies/classes taken for the study are quite distinct and separable,a class separability analysis using Jeffries–Matusita (JM) distancewas applied on the endmembers selected for the Hyperion andLandsat TM data. The Jeffries–Matusita (JM) distance is a commonlyapplied measure to quantify the degree of dissimilarity. The JM dis-tance value ranges from 0 (i.e. identical distributions) to 1.414 (i.e.complete dissimilarity).

4.6. Classification

SAM in general is less affected by illumination and topographiceffects however it has not been found well suited when appliedto complex feature spaces (Harken and Sugumaran, 2005; van derLinden et al., 2007). SVM which is a machine learning based pat-tern recognition technique is able to describe complex classes withmulti-modal distributions in the feature space. SVM has been foundadequate when classifying high dimensionality data with smalltraining sets (Pal and Mather, 2006).

Supervised classifications using SVM and SAM algorithms wereperformed on the Hyperion data. The classification was initiated bycollecting approximately 4–5 multiple endmembers per class fromthe image. The endmembers or training sites were carefully deter-mined based on the geographic position (latitude and longitude)of the classes taken during the field visit. The same endmemberswere used for classification of both the data, i.e. Hyperion andLandsat TM. For SVM classification the radial basis function kerneltype was selected with gamma value set to 0.34 for Hyperion and0.167 for Landsat TM and penalty parameter 100 and pyramid value0. In order to determine the optimal gamma and penalty values,commonly used grid search approach was used (Tso and Mather,2009). Classification was started with random parameter valueswhich are given inputs to the classifier to evaluate the t-fold cross-validation performance. Subsequently, the values of gamma andpenalty are increased or decreased, and the classification accuracyis re-evaluated until optimal parameters values are chosen.

SAM classification was performed only on Hyperion data. Thespectral similarity is determined by calculating the angle betweenreference spectra and satellite image spectra treating them as vec-tors in an n-dimensional space where ‘n’ equals the number ofbands used for the analysis. For SAM classification of Hyperion, thethreshold for different classes was selected after careful evaluationof different threshold angle value ranging from 0.05◦ to 0.3◦. The fol-lowing thresholds of maximum angle (radians) viz., chir pine (0.17),blue pine (0.10), cedar (0.10), grey oak (0.25), brown oak (0.20), fir(0.20) and grass (0.20) were iteratively decided. Selection of theparameters used in SAM classification was based on performinga number of trials of parameter combinations, using classificationaccuracy as a measure of quality (Petropoulos et al., 2012).

4.7. Accuracy assessment

Accuracies of SVM and SAM classified images were determinedin terms of error of omission and commission for different classes,and the overall classification accuracy using a contingency matrix

(Congalton, 1991). The overall accuracy provided the degree ofcloseness of classes mapped to the reference data collected usingGPS during field surveys. The kappa statistic was calculated to com-pare the true agreement between classes which actually occur on

144 R. George et al. / International Journal of Applied Earth Ob

Table 2Spectral similarity values between in situ and Hyperion image reflectance spectra.

S.No Common name Species botanical name SAM score

1 Chir pine Pinus roxburghii 0.6952 Blue pine Pinus wallichiana 0.8593 Grey oak Quercus leucotrichophora 0.8444 Brown oak Quercus semicarpifolia 0.857

to

5

5

raAtw

oifctrrstf

ttai

5

s

Fv

5 Fir Abies pindrow 0.8106 Cedar Cedrus deodara 0.880

he ground versus classified by the classifiers than those whichccur by chance.

. Results

.1. Evaluation of Hyperion spectra

Table 2 shows the spectral similarity scores between the in situeflectance spectra collected using SVC HR 1024 spectroradiometernd endmember spectra derived from Hyperion reflectance image.s indicated by the SAM score, there is a high similarity between

he in situ and Hyperion reflectance spectra. The score of chir pineas found lowest while it was highest for cedar.

Fig. 2 shows the visual comparison of in situ reflectance spectraf different species versus those derived from Hyperion reflectancemage. Though shapes of in situ and Hyperion reflectance spectraor different species are quite similar yet the amplitude of spectraollected using spectroradiometer is always higher as comparedo Hyperion reflectance spectra. The lower amplitude of Hyperioneflectance spectra is mainly due to its low signal to noise (SNR)atio compared to spectroradiometer spectra. The conformity inhapes of in situ and Hyperion reflectance spectra suggested thathe FLAASH atmospheric correction of the Hyperion dataset per-ormed well.

The evaluation of similarities between of shape of species spec-ra from Hyperion and field spectra provided the confidence thathe optimal bands identified through discriminant analysis wouldccount for the crown scale spectral variability among the speciesn the study area.

.2. Spectral separability

Tables 3 and 4 show the Jeffries–Matusita (JM) distance-basedpectral dissimilarity values for same endmembers derived from

ig. 2. Visual comparison of resampled field reflectance spectra of different speciesersus those from Hyperion.

servation and Geoinformation 28 (2014) 140–149

Hyperion and Landsat TM image. The JM value derived from end-members of Hyperion image ranged from 1.413 to 1.414 whichindicates high degree of spectral separability among the endmem-bers of different species classes. While the JM value derived fromthe endmembers from Landsat TM ranged from 1.034 to 1.414indicating a high to moderate separability but a lesser degree of sep-arability between species classes as compared to Hyperion image.

5.3. Spectral discriminant analysis

Wilk’s lambda (1978) ranked each band according to its ability toseparate the tree classes in the study area. Table 5 shows 29 optimalbands identified from the 160 continuum removed Hyperion bandsbased on the Wilk’s lambda values. The lesser the Wilk’s lambda,the greater is the separability between the vegetation classes. Of the29 bands selected in the 426 nm to 2400 nm wavelength region, 4bands were from visible, 8 bands from VNIR, and 17 bands fromSWIR regions.

5.4. Species discrimination and classification

Fig. 3 shows false colour composite of Hyperion, SVM and SAMclassified images of Hyperion, and SVM classified image of LandsatTM. In total six gregarious species, viz., chir pine, brown oak, greyoak, cedar, blue pine, fir besides mixed broadleaved and grasslands(comprising temperate and alpine) were mapped in the study area.In the lower elevation around 1600–1800 m, the southern aspectsare dominated by Chir pine along with grassy patches, whereas onthe northern aspect are occupied by grey oak and cedar in Tyasforest block. In the still higher altitude at 1700–2100 mts, in thesouthern altitude chir pine is replaced by blue pine and in thenorthern altitude, cedar, fir and mixed broadleaved species. Thesummit of Kedarkanta peak (3650 mts) is covered by snow andgrassland surrounds it. On the northern part of Kedarkanta peak,fir, mixed broadleaved and brown oak dominates. The distributionof the species clearly reveals the influence of elevation and aspecton species distribution.

The error matrices for accuracy assessment are shown in Table 6.Based on SVM classification of Landsat TM, the overall accuracyof 69.62% (kappa statistic 0.65) was achieved. For all the classesexcept chir pine (83.33%), the user’s classification accuracies werelow (<70%). Intermixing of broadleaved forests was found with fir,brown and grey oaks.

The species classification performed using SAM classifier onHyperion bands was better (overall accuracy 74.68%, kappa statis-tic 0.70) than SVM classification of Landsat TM. The improvementin classification accuracies was observed for the classes suchbrown oak and blue pine. However, the classification accuraciesof broadleaved, fir and cedar have decreased. Cedar showed inter-mixing with grey oak. Broadleaved showed intermixing with fir.

Among all the classifications (SVM on Landsat TM, SAM on Hype-rion and SVM on Hyperion), SVM on Hyperion produced the bestclassification (overall accuracy 82.27%, kappa statistic 0.79). Theuser accuracies of classes’ fir, blue pine and broadleaved classifiedusing SAM improved from 66.66%, 72.72%, and 72.72% to 72.72%,76.92% and 80% respectively when classified using SVM. Over-all SVM gave higher user’s accuracies (>80%) for classes such aschir pine, blue pine, and cedar. However, the user’s accuracies forremaining classes (grey oak, brown oak, broadleaved) was also bet-ter (>75%) except fir (72.72%) which has the lowest user’s accuracy.

6. Discussion

The Himalaya is one of the global biodiversity hotspots. Conser-vation and sustainable management of biodiversity in the Himalayais a priority agenda. Mapping and monitoring tree species location

R. George et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 140–149 145

Table 3Jeffries–Matusita (JM) distance values for Hyperion image training samples.

Grey oak Chir pine Cedar Fir Grass Blue pine Brown oak

Grey oak – 1.414 1.413 1.414 1.413 1.414 1.414Chir pine 1.414 – 1.413 1.414 1.414 1.414 1.414Cedar 1.413 1.414 – 1.414 1.414 1.414 1.414Fir 1.414 1.414 1.414 – 1.414 1.414 1.414Grass 1.413 1.414 1.414 1.414 – 1.414 1.414Blue pine 1.414 1.414 1.414 1.414 1.414 – 1.414Brown oak 1.414 1.414 1.414 1.414 1.414 1.414 –

Table 4Jeffries–Matusita (JM) distance values for Landsat TM image training samples.

Grey oak Chir pine Cedar Fir Grass Blue pine Brown oak

Grey oak – 1.391 1.382 1.413 1.414 1.406 1.376Chir pine 1.391 – 1.412 1.366 1.389 1.215 1.221Cedar 1.382 1.412 – 1.374 1.414 1.400 1.373Fir 1.413 1.366 1.374 – 1.395 1.146 1.195

aolopd

iin

Fc

Grass 1.414 1.389 1.414

Blue pine 1.406 1.215 1.400

Brown oak 1.376 1.221 1.373

nd habitats is essential for efficient planning and managementf biodiversity of the region. Hitherto mapping at species level ateast all the gregariously occurring species populations or patchesf sufficiently large sizes in the western Himalayan region was notossible through moderate resolution broadband remote sensingata.

In this study, an attempt has been made to generate anmproved species classification of the gregarious tree speciesn western Himalaya. The study has explored the potential ofarrow bands of Hyperion data and advanced hyperspectral data

ig. 3. FCC (NIR: 853, Red: 671, Green: 508), SVM and SAM classified images of Hyperionolour in this figure legend, the reader is referred to the web version of this article.)

1.395 – 1.400 1.3931.146 1.400 – 1.0341.195 1.393 1.034 –

classification techniques. Study has demonstrated the advantagesof stepwise discriminant analysis (Wilk’s lambda test statistic) foridentification of bands important for species discrimination. It isevident from Table 7 that the narrow bands identified in this studyare sensitive to various biochemical constituents found in plantspecies have therefore led to better interspecies separability in the

study area. Most of the identified bands belonged to NIR and SWIRregion of electromagnetic spectrum except the far end of SWIR. Inthe far end of SWIR, most of Hyperion bands are non-calibrated.In addition, bands in visible region appeared to exert a key role

, and SVM classified image of Landsat TM. (For interpretation of the references to

146

R.

George

et al.

/ International

Journal of

Applied

Earth O

bservation and

Geoinform

ation 28

(2014) 140–149

Table 5Wilk’s lambda values of 29 optimal wavebands.

Wavelength (nm) Wilk’slambda

701 0.105457 701 0.030457 701 732 0.009457 671 701 732 0.003457 671 701 732 1729 0.001457 508 671 701 732 1729 0.000457 508 569 671 701 732 1729 0.000457 508 569 671 701 732 853 1729 0.000457 490 508 569 671 701 732 853 1729 0.000457 490 508 569 671 701 732 853 932 1729 0.000457 490 508 569 671 701 732 853 932 972 1729 0.000457 490 508 569 671 701 732 853 932 972 993 1729 0.000457 490 508 569 671 701 732 853 932 972 993 1003 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1497 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1497 1628 1729 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1497 1628 1729 1810 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1497 1628 1729 1810 2032 0.000457 490 508 569 630 671 701 732 853 932 972 993 1003 1114 1154 1184 1205 1215 1316 1346 1457 1477 1487 1497 1628 1729 1810 2032 2052 0.000

R. George et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 140–149 147

Table 6Accuracy assessment results (i) SVM on 29 band Hyperion, (ii) SAM on 29 band Hyperion, and (iii) SVM on Landsat TM.

Chir pine Blue pine Grey oak Brown oak Fir Cedar Broadleaved Grassland Total UA

(i)Chir pine 11 0 1 0 0 1 0 0 13 84.61Blue pine 1 8 0 0 1 0 0 0 10 80Grey oak 0 0 10 2 0 0 1 0 13 76.92Brown oak 0 0 1 7 0 0 0 0 8 87.5Fir 0 2 0 0 8 0 1 0 11 72.72Cedar 0 0 0 0 0 11 0 0 11 100Broadleaved 0 0 0 1 2 0 10 0 13 76.92Total 12 10 12 10 11 12 12 10 79PA 91.66 80 83.33 70 72.72 91.66 83.33 100Overall accuracy: 82.27%, kappa statistics: 0.79

(ii)Chir pine 11 0 1 0 0 1 0 0 13 84.61Blue pine 1 8 0 0 1 1 0 0 11 72.72Grey oak 0 0 10 1 0 3 1 0 15 66.66Brown oak 0 0 1 6 0 0 0 0 7 85.7Fir 0 2 0 0 8 0 2 0 12 66.66Cedar 0 0 0 2 0 7 0 0 9 77.77Broadleaved 0 0 0 1 2 0 9 0 12 72.72Total 12 10 12 10 11 12 12 10 79PA 91.66 80 83.33 60 72.72 58.33 75 100Overall Accuracy: 74.68%, Kappa Statistics: 0.70

(iii)Chir pine 10 0 1 0 0 1 0 0 12 83.33Blue pine 1 7 0 1 1 0 0 0 10 70Grey oak 0 0 10 1 0 3 1 0 15 66.66Brown oak 0 0 1 5 0 0 1 0 7 71.42Fir 1 2 0 0 9 0 4 0 16 56.25Cedar 0 0 0 2 0 8 0 0 10 80Broadleaved 0 1 0 1 1 0 6 0 9 66.66Total 12 10 12 10 11 12 12 10 79PA 83.33 70 83.33 50 81.81 66.66 50 100Overall Accuracy: 69.62%, Kappa Statistics: 0.65

Table 7The selected wavebands and their importance for vegetation.

S.No Wavelength (nm) Absorbing compound Reference

1. 457 Chlorophyll b Curran et al. (1991) and Kumar et al. (2001)2. 490 Nitrogen Asner (2008) and Vyas et al. (2011)3. 508 Nitrogen Asner (2008) and Vyas et al. (2011)4. 569 Pigments (Anthocyanins, Chlorophyll), Nitrogen Thenkabail (2002)5. 630 Chlorophyll b Kumar et al. (2001) and Vyas et al. (2011)6. 671 Chlorophyll (Red 2) Thenkabail (2002)7. 701

Chlorophyll + nitrogenClevers and Bucker (1991), Curran et al. (1991)and Filella and Penuelas (1994)8. 732

9. 853 Biophysical\biochemical quantities, Heavy metal stress Thenkabail (2002) and Cochrane (2000)10. 932 Oil Kumar et al. (2001) and Cochrane (2000)11. 972 NIR-moisture absorption Thenkabail (2002)12. 993 Starch Kumar et al. (2001)13. 1003 Water content Vyas et al. (2011)14. 1114 Lignin Serrano et al. (2002)15. 1154 Moisture absorption –16. 1184 Moisture absorption –17. 1205 Water, cellulose, starch, lignin Kumar et al. (2001)18. 1215 Moisture absorption Thenkabail et al. (2004)19. 1316 Moisture absorption Sobhan (2007)20. 1346 Nitrogen Curran (1989) and Kumar et al. (2001)21. 1457 Lignin Kumar et al. (2001)22. 1477 Cellulose Kumar et al. (2001) and Vyas et al. (2011)23. 1487 Nitrogen Curran (1989) and Kumar et al. (2001)24. 1497 Protein Curran (1989) and Kumar et al. (2001)25. 1628 Moisture absorption Sobhan (2007)26. 1729 Protein –27. 1810 Cellulose Schmidt and Skidmore (2003)28. 2032 Moisture absorption Thenkabail et al. (2004)29. 2052 Protein Kumar et al. (2001) and Sobhan (2007)

1 rth Ob

irwsw

pTHbetontsctaE

distcsvctip

tascedHas

7

tihcThhoabssofidpos

48 R. George et al. / International Journal of Applied Ea

n discrimination of species. The narrow bands selected in visibleegion have ability to sense to leaf pigments like chlorophyll,hile those chosen in NIR regions are sensitive to leaf and canopy

tructure, and the bands chosen in SWIR region are sensitive toater content, nitrogen, cellulose, starch and sugar.

Study has shown that classification of narrow band data canrovide better species classification accuracy than broadband data.he classification accuracy achieved using SVM classification ofyperion bands was much better than Landsat TM bands both forroadleaved and conifer species. Further, SVM was found moreffective in discrimination of intermixed patches of gregariousree species from several narrow bands compared to SAM. SVMffered higher classification accuracy despite its lower signal tooise (SNR) compared to Landsat TM. Previous studies have shownhe promise for improving species discrimination through hyper-pectral remote sensing. Goodenough et al. (2003) have comparedlassification results from Hyperion, ALI, and the Landsat ETM+ forhe Greater Victoria Watershed, Canada and obtained 92.9%, 84.8%nd 75.0% overall classification accuracies for Hyperion, ALI andTM+ respectively.

In this study better species discrimination was achieved alsoue to selection of right time satellite data. During April month

ntraspecific variability in the study area is found minimum whilepectral differences among the species are found at maximum. Allree species in the study area were evergreen in nature excepthir pine which has semi-deciduous character. Since all the treepecies were in leafing therefore low intraspecific variability pro-ided better contrast for species discrimination. However, lowanopy density as in case of fir (ranged from 30 to 50%) might exertshe influence of background soil and undergrowth on the acquiredmage which leads to high intraspecific variability. This can causeroblem in spectral segregation.

The shadow in the remotely sensed satellite data of moun-ain area is a common problem. Shadow reduces the classificationccuracies of target tree species. In this study, it was found thatpecies classification using SVM was less affected by shadowompared to SAM. Further, SVM was able to tackle shadowffects more efficiently in narrow bands data than broad bandsata as indicated by the higher accuracy achieved for fir usingyperion bands. In general, shadow affected the classificationccuracies of conifers more than broadleaved on the northernlopes.

. Conclusions

The discrimination and classification of broadleaved and coniferree species in western Himalayas is challenging with the lim-ted spectral information from multispectral data. Present workas confirmed the potential utility of hyperspectral data to dis-riminate tree species in mountainous areas at larger scales.he study emphasizes the careful selection of wavebands fromyperspectral data and classification algorithm for producingigh accuracy forest tree species maps. The optimal selectionf wavebands can be attempted using step-wise discriminantnalysis which can identify suitable and considerably lower num-er of wavebands meaningful for discrimination of range ofpecies occurring in the particular area. SVM produced a betterpecies classification (82.88% overall accuracy) than SAM (75.12%verall accuracy). A significant improvement in species classi-cation was achieved using Hyperion bands over Landsat TM

ata (69.43% overall accuracy). The non-parametric classifier hasroven superior in handling information from a large numberf bands for effective classification of intermixed patches of treepecies.

servation and Geoinformation 28 (2014) 140–149

Acknowledgements

The authors are thankful to Dr. Y. V. N. Krishna Murthy, Director,Centre for Space Science and Technology Education in the Asia andthe Pacific (CSSTEAP). The first author thanks Dr. Sarnam Singh andDr. Yogesh Kant from CSSTEAP for support during the study. We arethankful to Dr. D. Vyas for valuable suggestions on data processing.Thanks are also due to Dr. Gajendra Rawat for assistance in fielddata collection.

References

Asner, G.P., 2008. Hyperspectral remote sensing of canopy chemistry, physiology andbiodiversity in tropical rain forests. In: Kalascska, M., Arturo Sanchez-Azofeifa,G. (Eds.), Hyperspectral Remote Sensing of Tropical and Sub-Tropical Forests.CRC press Taylor and Francis group, Boca Raton, London, New York, pp.161-196.

Asner, G.P., Heidebrecht, K.B., 2002. Spectral unmixing of vegetation, soil anddrycarbon cover in arid regions: comparing multispectral and hyper-spectralobservations. International Journal of Remote Sensing 23, 3939–3958.

Champion, H.G., Seth, S.K., 1968. A Revised Survey of the Forest Types of India. Govt.of India Press, New Delhi.

Chan, J.C.W., Palinckx, D., 2008. Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mappingusing airborne hyperspectral imagery. Remote Sensing of Environment 112,2999–3011.

Clark, M.L., Roberts, D.A., Clark, D.B., 2005. Hyperspectral discrimination of tropicalrain forest tree species at leaf to crown scales. Remote Sensing of Environment96, 375–398.

Clevers, J., Bucker, C., 1991. Feasibility of the red edge index for the detection ofnitrogen deficiency. In: Proceedings of the 5th International Colloquium onPhysical Measurement and Signatures in Remote Sensing, Courcheval, France,pp. 165–168.

Congalton, R.G., 1991. A review of assessing the accuracy of classifications ofremotely sensed data. Remote Sensing of Environment 37, 35–46.

Cochrane, M.A., 2000. Using vegetation reflectance variability for species level clas-sification of hyperspectral data. International Journal of Remote Sensing 21 (10),2075–2087.

Curran, P.J., 1989. Remote sensing of foliar chemistry. Remote Sensing of Environ-ment 30 (3), 271–278.

Curran, P.J., Dungan, J.L., Macler, B.A., Plummer, S.E., 1991. The effect of a red leafpigment on the relationship between red edge and chlorophyll concentration.Remote Sensing of Environment 35, 69–76.

Dalponte, M., Bruzzone, L., Vescovo, L., Gianelle, D., 2009. The role of spectral reso-lution and classifier complexity in the analysis of hyperspectral images of forestareas. Remote Sensing of Environment 113, 2345–2355.

Du, Q., Ren, H., 2003. Real-time constrained linear discriminant analysis to targetdetection and classification in hyperspectral imagery. Pattern Recognition 36,1–12.

Erbek, F.S., Ozkan, C., Taberner, M., 2004. Comparison of maximum likelihood clas-sification method with supervised artificial neural network algorithms for landuse activities. International Journal of Remote Sensing 25, 1733–1748.

Filella, I., Penuelas, J., 1994. The red edge position and shape as indicators of plantchlorophyll content, biomass and hydric status. International Journal of RemoteSensing 15, 1459–1470.

Foody, G.M., 2004. Supervised image classification by MLP and RBF neural networkswith and without an exhaustively defined set of classes. International Journal ofRemote Sensing 25, 3091–3104.

Green, P.E., Caroll, J.D., 1978. Analyzing Multivariate Data. The Dryden Press, Illinois.Goodenough, D.G., Dyk, A., Niemann, O.K., Pearlman, J.S., Chen, H., Han, T., Murdoch,

M., West, C., 2003. Processing Hyperion and ALI for forest classification. IEEETransaction on Geoscience and Remote Sensing 41, 1321–1331.

Gomez-Chova, L., Calpe, J., Camps-Valls, G., Martín, J., Soria, E., Vila, J., Alonso-Chorda,L., Moreno, J., 2003. Feature selection of hyperspectral data through local cor-relation and SFFS for crop classification. In: IEEE International Geoscience andRemote Sensing Symposium, IGARSS03, vol. I, pp. 555–557.

Harken, J., Sugumaran, R., 2005. Classification of Iowa Wetlands using an airbornehyperspectral image: a comparison of Spectral Angle Mapper (SAM) Classifierand an Object-Oriented (OO) approach. Canadian Journal of Remote Sensing 31(2), 1–8.

Jain, N., Ray, S.S., Singh, J.P., Panigrahy, S., 2007. Use of hyperspectral data to assessthe effects of different nitrogen applications on a potato crop. Precision Agricul-ture 8, 225–239.

Kent, M., Coker, P., 1992. Vegetation Description and Analysis. A Practical Approach.John Wiley and Sons, London.

Kruse, F.A., Lefkoff, A.B., Boardman, J.B., Heidebrecht, K.B., Shapiro, A.T., Barloon,P.J., Goetz, A.F.H., 1993. The Spectral Image Processing System (SIPS) – interac-

tive visualization and analysis of imaging spectrometer data. Remote Sensing ofEnvironment 44, 145–163.

Kumar, L., Schmidt, K.S., Dury, S., Skidmore, A.K., 2001. Imaging spectrometry andvegetation science. In: van der Meer, F., de Jong, S.M. (Eds.), Imaging Spectrom-etry. Kluwer Academic publishers, Dordrecht, The Netherlands.

rth Ob

L

L

L

M

M

M

N

O

P

P

P

P

P

R. George et al. / International Journal of Applied Ea

andgrebe, D.A., 2003. Signal Theory Methods in Multispectral Remote Sensing. JohnWiley and Sons, Hoboken, NJ, USA.

awrence, R., Bunn, A., Powell, S., Zmabon, M., 2004. Classification of remotely sensedimagery using stochastic gradient boosting as a refinement of classification tree-analysis. Remote Sensing of Environment 90, 331–336.

ucas, R., Bunting, P., Paterson, Chisholm, L., 2008. Classification of Australian forestcommunities using aerial photography, CASI and HyMap data. Remote Sensingof Environment 110, 2088–2103.

elgani, F., Bruzzone, L., 2004. Classification of hyperspectral remote sensing imageswith support vector machines. IEEE Transactions on Geoscience and RemoteSensing 42, 1778–1790.

ountrakis, G., Im, J., Ogole, C., 2011. Support vector machines in remote sensing: areview. ISPRS Journal of Photogrammetry and Remote Sensing 66, 247–259.

yint, S.W., 2001. A robust texture analysis and classification approach for urbanland-use and land-cover feature discrimination. Geocarto International 16,27–38.

eville, R.A., Levesque, J., Staene, K., Nadeau, C., Hauff, P., Borstad, G.A., 2003. Spec-tral unmixing of hyperspectral imagery for mineral exploration: comparison ofresults from SFSI and AVIRIS. Canadian Journal of Remote Sensing 29, 99–110.

kin, G.S., Roberts, D.A., Murray, B., Okin, W.J., 2001. Practical limits on hyperspectralvegetation discrimination in arid and semiarid environments. Remote Sensingof Environment 77, 212–225.

al, M., Mather, P.M., 2003. An assessment of the effectiveness of decision tree meth-ods for land cover classification. Remote Sensing of Environment 86, 554–565.

al, M., Mather, P.M., 2006. Some issues in the classification of DAIS hyperspectraldata. International Journal of Remote Sensing 27, 2895–2916.

etropoulos, G.P., Kalaitzidis, C., Vadrevu, K.P., 2012. Support Vector Machine andobject based classification for obtaining landuse/landcover cartography fromHyperion hyperspectral imagery. Computers and Geosciences 41, 99–107.

latt, R.V., Goetz, A.F.H., 2004. A comparison of AVIRIS and Landsat for land use clas-sification at the urban fringe. Photogrammetric Engineering and Remote Sensing

70, 813–819.

laza, A., Benediktsson, J.A., Boardman, J.W., Brazile, J., Bruzzone, L., Camps-Valls,G., Chanussot, J., Gamba, P., Fauvel, M., Gualtieri, A., Marconcini, M., Yoilton,J.C., Trianni, G., 2009. Recent advances in techniques for hyperspectral imageprocessing. Remote Sensing of Environment 113, 110–122.

servation and Geoinformation 28 (2014) 140–149 149

Rashed, T., Weeks, J.R., Roberts, D., Rogan, J., Powell, R., 2003. Measuring the physicalcomposition of urban morphology using multiple endmember spectral mixtureanalysis. Photogrammetric Engineering and Remote Sensing 69, 1011–1020.

Ray, S.S., Jain, N., Miglani, A., Singh, J.P., Singh, A.K., Panigrahy, S., Parihar, J.S., 2010.Defining optimum spectral narrow bands and bandwidths for agricultural appli-cations. Current Science 98, 1365–1369.

Schmidt, K.S., Skidmore, A.K., 2003. Spectral discrimination of vegetation types in acoastal wetland. Remote Sensing of Environment 85, 92–108.

Serrano, L., Penuelas, J., Ustin, S.L., 2002. Remote sensing of nitrogen and lignin inMediterranean vegetation from AVIRIS data: decomposing biochemical fromstructural signals. Remote Sensing of Environment 81, 355–364.

Sobhan, I.Md., 2007. Species discrimination from a hyperspectral perspective. Ph.D.thesis. ITC, The Netherlands.

Soman, C.R., Ajay, V., Loganathan, R., 2009. Machine Learning with SVM and OtherKernel Methods. Prentice-Hall, India.

Thenkabail, P.S., Enclona, E.A., Ashton, M.S., Legg, C., Dieu, M.J.D., 2004. Hyperion,IKONOS, ALI, and ETM+ sensors in the study of African rainforests. RemoteSensing of Environment 90, 23–43.

Thenkabail, P.S., 2002. Optimal hyperspectral narrow bands for discriminating agri-cultural crops. Remote Sensing Reviews 20, 257–291.

Tso, B., Mather, P.M., 2009. Classification Methods for Remotely Sensed Data, 2nded. CRC Press, Boca Raton.

Van Aardt, J.A.N., Wynne, R.H., 2001. Spectral separability among six southern treespecies. Photogrammetric Engineering and Remote Sensing 67, 1367–1375.

van der Linden, S., Waske, B., Hostert, P., 2007. Towards an optimized use of thespectral angle space. In: Imaging Spectrometry: Innovation in EnvironmentalResearch: Proceeding of the 5th workshop of the EARSel Special Interest Groupon Imaging Spectroscopy, 23–25 April 2007, Bruges, Belgium. EARSel Secretariat,Hannover, Germany.

Vyas, D., Krishnayya, N.S.R., Manjunath, K.R., Ray, S.S., Panigrahy, S., 2011. Evalu-ation of classifiers for processing Hyperion (EO-1) data of tropical vegetation.

International Journal of Applied Earth Observation and Geoinformation 13,228–235.

Waske, B., Benediktsson, J., Árnason, K., Sveinsson, J., 2009. Mapping of hyperspectralAVIRIS data using machine-learning algorithms. Canadian Journal of RemoteSensing 35, S106–S116.