ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65

Contents lists available at SciVerse ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier .com/ locate/ isprs jprs

Discriminating indicator grass species for rangeland degradation assessmentusing hyperspectral data resampled to AISA Eagle resolution

Khalid Mansour a,⇑, Onisimo Mutanga b, Terry Everson c, Elhadi Adam b

a University of KwaZulu-Natal, School of Applied Environmental Sciences, Centre for Environment, Agriculture & Development (CEAD), P/Bag X01, Scottsville, Pietermaritzburg 3209,South Africab University of KwaZulu-Natal, School of Applied Environmental Sciences, Geography Department, P/Bag X01, Scottsville, Pietermaritzburg 3209, South Africac University of KwaZulu-Natal, School of Biological and Conservation Sciences, P/Bag X01, Scottsville, Pietermaritzburg 3209, South Africa

a r t i c l e i n f o

Article history:Received 22 February 2011Received in revised form 31 January 2012Accepted 13 March 2012Available online 28 April 2012

Keywords:Rangeland degradationRandom forestIndicator speciesField spectrometer measurementsVariable selection

0924-2716/$ - see front matter � 2012 Internationalhttp://dx.doi.org/10.1016/j.isprsjprs.2012.03.006

⇑ Corresponding author.E-mail address: wadmanssour@gmail.com (K. Man

a b s t r a c t

The development of techniques to estimate and map increaser grass species is critical for better under-standing the condition of the rangeland and levels of rangeland degradation. This paper investigateswhether canopy reflectance spectra, resampled to AISA Eagle resolution can discriminate among fourincreaser species representing different levels of rangeland degradation. Canopy spectral measurementswere taken from the four indicator species: Hyparrhenia hirta (HH), Eragrostis curvula (EC), Sporobolusafricanus (SA), and Aristida diffusa (AD). The random forest algorithm and a forward variable selectiontechnique were used to identify optimal wavelengths for discriminating the species. Results revealed thatthe optimal number of wavelengths (n = 8) that yielded the lowest OOB error (11.36%) in discriminatingamong the four increaser species are located in 966.7, 877.6, 691.9, 718.7, 902.7, 854.8, 674.1 and703 nm. These wavelengths are located in the visible, red-edge and near-infrared regions of the electro-magnetic spectrum. The random forest algorithm can accurately discriminate species with an overallaccuracy of 88.64% and a KHAT value of 0.85. The study demonstrated the possibility to upscale themethod to airborne sensors such as AISA Eagle for mapping indicator species of rangeland degradation.A rotational grazing management plan should be considered as a way to create sustainable rangelandmanagement in degraded areas.� 2012 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS) Published by Elsevier

B.V. All rights reserved.

1. Introduction

Rangeland degradation is defined as the reduction or temporaryloss of the biological and economic productivity of grasslands in arid,semi-arid, and dry sub-humid areas (UNCCD, 1995). Currently,rangeland degradation has been identified as one of the most seriousglobal environmental issues (Wessels et al., 2007). Approximatelyover 250 million people in over 100 countries are directly affectedby rangeland degradation (Adger et al., 2000; Wessels et al., 2007).In South Africa, rangeland degradation is believed to be one of themost severe and widespread environmental problems (Hoffmanand Todd, 2000; Wessels et al., 2004). A total of 4.8% (5.8 millionha) of the South African land has been identified as degraded as indi-cated by lower vegetation cover compared to the surrounding areas(Thompson, 1996; Wessels et al., 2004). The greatest areas of exten-sively degraded land coincide with communal lands and rangelandswhere a considerable population of South Africa and livestock lives(Hoffman and Todd, 2000; Reid and Vogel, 2006). Many South

Society for Photogrammetry and R

sour).

African studies on rangeland degradation have been concentratedon commercial areas (Palmer and van Rooyen, 1998; Shackletonet al., 2005). However, the communal areas have not yet receivedthe same level of attention that has been apparent in the commercialareas (Hoffman and Todd, 2000; Wessels et al., 2004). The continuedrangeland degradation represents a significant threat to the live-stock and biodiversity (Lorent et al., 2008). Therefore, there is a needfor planning strategies to map and monitor rangeland degradation atdifferent scales using consistent, repeatable, and spatially explicitmeasures (Prince et al., 2009; Ravi et al., 2010). These planning strat-egies for sustainable land management require techniques that caneffectively reveal the spatial extent, magnitude, and temporalbehavior of the lands (Prince et al., 2009; Ravi et al., 2010; VanLynden and Mantel, 2001). Remote sensing techniques provide anefficient cost-effective means to assess and map rangelanddegradation (Ustin et al., 2009). However the use of remote sensingtechniques in mapping rangeland degradation requires simple indi-cators that allow combining ground-based methods with remotelysensed data (Pyke et al., 2002). Several indicators have been sug-gested for mapping rangeland degradation such as soil organic mat-ter (Wang et al., 2010), vegetation production (Wessels et al., 2008),

emote Sensing, Inc. (ISPRS) Published by Elsevier B.V. All rights reserved.

K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65 57

natural and semi-natural vegetation communities (Hill et al., 2008).The limitation of these studies is that they have mainly been focusedon binary maps that identify the degraded and non-degraded areas.Although these methods can allow drawing the line between thetwo classes, they do not allow identifying different levels of range-land degradation using indicators that can easily and directly bedetected and monitored. Such indictor could be vegetation species.This is because certain vegetation species are well adapted tospecific growth conditions and their quality and quantity reduceor increase according to the change on the growth conditions(Nordberg and Allard, 2002; Van oudtshoorn, 1992).

In South Africa, grassland species have been classified into twogroups of increaser and decreaser species based on changes in theirrelative abundance in the presence or absence of grazing whichindicates the condition of the rangeland (Vesk and Westoby,2001). Increaser species are species that increase their relativeabundances through grazing or under-utilization, and thereforeindicate the poor condition of the rangeland (Vesk and Westoby,2001; Van Oudtshoorn, 1992). Increaser species have been classi-fied into three types, namely, increaser I, increaser II, and increaserIII (du Toit, 2009; Oluwole et al., 2008; Trollope, 1990). Increaser Ispecies such as Hyparrhenia hirta increase in abundance with un-der-utilization and can be found in areas with low grazing capacity(e.g. conserved areas), while increaser II species increase in abun-dance when the rangeland is over-utilized (e.g. Eragrostis curvulaand Sporobolus africanus), and increaser III species (e.g. Aristidadiffusa) increase in relative abundance in rangeland that is selec-tively grazed (du Toit, 2009; Oluwole et al., 2008; Trollope, 1990;Van oudtshoorn, 1992). A gradient of degradation has been classi-fied to range from severe with a high relative abundance of increa-ser I, increaser II, and increaser III species to non-degradedrangeland with a high abundance of decreaser species. The four in-creaser species I, IIa, IIb, and III were considered in this study sincethese are the only dominant species in the study area that havebeen identified as indicators of rangeland degradation. Therefore,the relative abundance and distribution of these different increaserspecies can be used to classify rangeland condition into moderate(increaser I), poor (increaser II) and highly degraded (increaserIII), thereby indicating the gradient of rangeland degradation.

Up-to-date spatial information about increaser species is essen-tial for classifying rangeland condition. To our knowledge, no at-tempt has yet been made to discriminate increaser species withremote sensing as indicators of the different levels of rangelanddegradation.

Traditionally, mapping vegetation species generally needsintensive fieldwork, including identification of species characteris-tics and the visual estimation of species percentage all of which arecostly and time-consuming and sometimes impossible to accom-plish due to poor accessibility (Adam et al., 2009; Muchoney andHaack, 1994). On the other hand, remote sensing techniques offeran economic and effective technique, producing timely andaccurate information for mapping vegetation species (Ustin et al.,2009).

Hyperspectral remote sensing, in particular, is developing as amore in-depth means of investigating spatial, temporal, and spec-tral discrimination of vegetation species quantity and quality (Us-tin et al., 2009). This is due to its use of many narrow andcontiguous spectral bands of less than 10 nm. These bands allowthe detection of vegetation at species levels which are otherwisemasked by broad bands of multispectral satellites such SPOT (Ku-mar et al., 2001; Mutanga and Skidmore, 2004; Mutanga et al.,2005). Hyperspectral remote sensing data are acquired usingspaceborne, airborne sensors and a hand-held spectrometer (Adamet al., 2009). At the moment, hyperspectral remote sensing has notreached operational level at a wider scale due to the costs ofimages and the small areal extent covered by airborne images.

However, research on the behavior of indicator vegetation speciesusing field spectroscopic data is an important step towards under-standing the critical bands, absorption features and curves that canbe targeted for building operational sensors to reveal the behav-ioral patterns of rangeland degradation. Processing hyperspectralremote sensing data is challenging due to the high dimensionality,overfitting when applying statistical methods, an excessive de-mand for sufficient field samples, and high cost (Bajcsy and Groves,2004; Vaiphasa et al., 2007). Therefore, identifying the optimal andpowerful wavelengths using variable selection methods withoutlosing any important information is a pre-requisite in hyperspec-tral remote sensing application (Adam and Mutanga, 2009; Bajcsyand Groves, 2004; Vaiphasa et al., 2007). This method is done, notonly to reduce the amount of variables to simplify the model, butalso to determine which explanatory variables are most suitablein classifying increaser species. Different statistical techniqueshave been used to identify the optimal wavelengths such as dis-criminant analysis, canonical variate analysis, classification trees,support vector machines; and principal component analysis (Adamand Mutanga, 2009; Cochrane, 2000; Mutanga and Skidmore,2004).

Recently, the random forest algorithm which was developed byBreiman (2001), has been successfully used as a variable selectionand classification algorithm for hyperspectral data (Adam et al.,2009; Ismail, 2010; Lawrence et al., 2006). Random forest is a treeensemble algorithm that uses a bagging, i.e., bootstrap aggregation,ensemble procedure to build multiple individual decision treesthat are provided to be diverse by the use of random samples de-rived from the training data set (Breiman, 2001). The training datais sampled to create an in-bag partition to construct the tree (2/3 ofthe training data), and a smaller out-of-bag partition (1/3 of thetraining data set) to validate the performance of each constructedtree (Özçift, 2011). The multiple trees then vote by majority on cor-rect classification.

The objectives of this study were to investigate the use of ran-dom forest algorithm to identify crucial wavelengths that are mostsensitive in discriminating four indicator species (H. hirta, E. curv-ula, S. africanus and A. diffusa) for different levels of rangeland deg-radation in the Okhombe, South Africa. We also sought toinvestigate whether canopy reflectance spectra, resampled to AISAEagle spectral resolution, could be used to discriminate amongthese four species using random forest algorithms.

2. Materials and methods

2.1. Study area

Okhombe is a communal grazing land in the province of KwaZ-ulu-Natal, South Africa (Fig. 1), located within 10–20 km of thenorth-eastern border of Lesotho, between latitudes 28�300S and30�300S and longitudes 28�300E and 29�300E. The average altitudefor the site varies from about 1200 to 3350 m with an averageair temperature of 11.5 to 16 �C in the summer months (Octo-ber–March). In winter (June and July), the mean monthly temper-ature reaches 5 �C with frost and snow occurring almost everywinter (Temme, 2008). The mean annual rainfall of the area isabout 800–1000 mm, and it receives about 82% of this rainfall inthe summer months (Dollar and Goudy, 1999). The geology ofthe area is a mixture of mudstone, sandstone, tillite, ampholite,and basalt prevailing on slopes and plateaux. Soils in the area areGriffin and Hutton (Oxisol), Clovelly (Alfisol), Mispah and Glenrosa(Ebtisol/Enceptisol) on slopes and plateaux. The study area is com-prehensively covered by grassland with some patches of forest,shrubs, and tree species. Vegetation communities are associatedwith three distinct altitudinal zones (O’Connor and Bredenkamp,1997). These zones are the river valleys (1250–1800 m), the Little

Fig. 1. Location of study area in KwaZulu-Natal (KZN) province of South Africa.

58 K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65

Berg (1800–2500 m), and the summit plateaux (2500–3350 m).The common vegetation species in these zones are: Hyparrheniaspecies, Eragrostis species, Aristida species, Themeda triandra, Digi-taria species, Panicum species, Monocymbium ceresiiforme, Sporobo-lus species, and Miscanthus capense (O’Connor and Bredenkamp,1997). Natural factors such as climate, lithology, and geomorphol-ogy, and man induced factors such as population increase, over-grazing by livestock, out of season burning and deforestationhave accelerated vegetation degradation, for example the disap-pearance of palatable species and an increase in unpalatable grass

species (Critchley and Netshikovhela, 1998; Everson and Tainton,2007; Hoffman et al., 1999).

2.2. Field data collection

2.2.1. The identification of increaser grass speciesIntensive field work was conducted to identify the grass species

that are associated with rangeland degradation in the study area(Fig. 2). Four indicator grass species were then selected based ontheir high relative abundance. These species are H. hirta (HH),

Fig. 2. Visual indicators of rangeland degradation observed in Okhombe (A) cattle access routes (Increaser I), (B) sedimentation in streams (Increaser II), (C) gullies(Increaser III).

K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65 59

E. curvula (EC), S. africanus (SA), and A. diffusa (DA). These speciesrepresent increaser I, increaser II and increaser III categories (Table1).

2.2.2. Canopy spectral measurementsSpectral measurements at canopy level were taken from the

four increaser species (H. hirta,E. curvula, S. africanus, and A. diffusa)using the Analytical Spectral Devices (ASD) FieldSpec� 3 (ASD). Thespectral range of the ASD is 350–2500 nm with a resolution of1.4 nm in the 350–1000 nm range and 2.0 nm for the spectral re-gion 1000–2500 nm (ASD, 2005). Random points were generatedusing Hawth’s Analysis Tool (HAT) in ArcGIS 9.3 (Adam and Mutan-ga, 2009) and an existing land cover map of the study area devel-oped by the research group (Bangamwabo, 2009). A species plotwas defined to cover 3 m � 3 m, where the target species (n = 4)were more homogenous with high relative abundance of morethan 80% of the target species in each plot. A total of 75 plots weregenerated for each grass species (HH, EC, SA, and AD). A total of20–25 spectral measurements were then taken randomly in eachplot at nadir from 1.5 m using a 5� field of view (Table 2). Thisyielded a ground field of view of about 13 cm above the leaveson a clear sunny day of 21st of November between 11:00 am and2:30 pm local time (Greenwich Mean Time: GMT + 2). These spec-tral measurements from each plot (n = 20–25) were then averagedto represent the spectral reflectance of the vegetation plot(n = 308). The spectral measurements were then resampled tothe AISA Eagle spectral resolution using ENVI 4.3 image processingsoftware (Mutanga, 2005) (Fig. 3). AISA Eagle data has a 2 m spatialresolution and a spectral range from 393.2 to 994.1 nm (272 wave-lengths) at 2.04 to 2.29 nm spectral resolutions. The resampledAISA Eagle spectra were then used for subsequent analysis. Thedata set for each target species was then split randomly into70/30 training data set (n = 53) and test data set (n = 22), respec-tively (Ismail and Mutanga, 2011).

3. Data analysis

3.1. Measuring variable importance using the random forest algorithm(RF)

The random forest algorithm is a forest-based method devel-oped by Breiman (2001) to overcome the instability of traditionaltree-based methods. Breiman (2001) defined random forests algo-rithm as follows: random forest is consist of collection of multipledecision tree classifiers that defined as fhðx;HkÞ; k ¼ 1; �g. Where

Hk represents identically distributed random vectors and each treecasts a unit vote for the most popular class at input X.

Each decision tree in the forest is constructed by the followingsteps (Fig. 4):

(1) The number of trees (T) to be grown is selected.

(2) The number of variables (f) to split each node is chosen. Ifthe variables of the input data is denoted by F, then f < Fmust be satisfied. The subset of features f is kept constantduring the formation of forest.

(3) T number of trees (ntree) is grown with the followingcriteria:

(a) A bootstrap sample of size n is constructed set withreplacement and a sample of Sn is selected to grow a tree.(b) To grow a tree at each node, m features are selected ran-domly and they are used to find the best split.(c) Each tree is grown to maximum size without pruning.

To classify a sample X, (in our case the increaser species) amajority voting scheme is used to evaluate votes from each treein the forest.

The RF algorithm provides three independent variable impor-tance measures, specifically, the permutation accuracy importancemeasure, the Gini importance, and the number of times each vari-able is selected (Breiman, 2001). The permutation accuracy impor-tance measure, is considered to be the best measure in randomforests because of its capability to assess the variable importancethat relies on mean decreases in accuracy as measured using theout-of-bag (OOB) samples (Breiman, 2001). The OOB (out of bag)refers to the samples not included in the bootstrap iteration (Brei-man, 2001). The OOB error rate is computed by putting each OOBobservation down the corresponding classification tree from whichit was excluded. The error estimate is then computed as the mis-classified proportion of that OOB observation. The OOB error pro-duces a measure of the importance of the variables by comparinghow much the OOB error of estimate increases when a variableis permutated whilst all other variables are left unchanged (Archerand Kimes, 2008; Peters et al., 2007). The variables permutation isthe most reliable measure which computes variable importance asthe mean decrease in accuracy based on the OOB observations(Breiman, 2001). According to Ismail and Mutanga (2011) thereflectance values of each wavelength were randomly permuted,its original association with the response variable is broken. Thepermuted wavelength and the remaining non-permuted wave-lengths are then used to classify the response for the OOB observa-tion. It follows that the classification accuracy will decrease

Table 1Visual indicators of Okhombe rangeland degradation based on different increaser species.

Indicator species Common name General characteristics Grazing value* Visual indicators ofrangeland degradation

Degradation stage

Increaser I (HH) Thatching grass A relatively dense, perennial tuftedgrass. Spikelets are covered withwhite to grey hairs. Culms 300–1500 mm tall, leaf blade 1–4 mmwide. Flowers from September toMarch

5 Bare soil on cattle access routes,accumulations of soil around treesand fences, dust storm, muddy water

Poor

Increaser II (EC, SA) EC: Weeping lovegrass EC: densely perennial tufted grass,inflorescences are mostly an openpanicle. Spikelets are dark grey todark olive green, culms 300–1200 mm tall, leaf blade up to 4 mmwide. Flowers from August to June

3–5 Barren spot, Sandy layer on soilsurface, vetiver grass, damagedswales, Sedimentation in streams

Moderate

SA: Ratstail Dropseed SA: perennial tufted grass, longpanicle with a pointed tip, Culms280–1500 mm tall, leaf blade 1–4 mm wide. Flowers from October toApril

Increaser III (DA) Iron grass A tufted perennial grass. Leaves arehard, narrow and rolled.Inflorescences are a spare, expandedand open panicle. Culms 300–800 mm tall, leaf blade up to 2 mmwide. Flowers from November toApril

0 Bare soil, eroded slope, rills, gullies,exposed roots, Dongas, parentmaterial (stones)

High

* van Oudtshoorn (1992).

Table 2Species name, number of sample plots, and the total number of spectralmeasurements.

Species name Type code No. of plots Spectral measurements

Hyparrhenia hirta HH 75 1730Eragrostis curvula EC 75 1700Sporobolus africanus SA 75 1715Aristida diffusa AD 75 1780

60 K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65

substantially if the original wavelength was associated with theresponse variable. Thus, the classification accuracy differencebefore and after permuting the wavelengths can be applied as avariable importance measure (Breiman, 2001).

In this paper, the permutation of variables (mean decrease inaccuracy) to measure the importance of AISA Eagle wavelengthsin discriminating the increaser species was used (Breiman, 2001)as a ranking index to measure of the importance of the variableand thereafter identifying the wavelengths with relatively largeimportance in the classification process (Archer and Kimes, 2008;Díaz-Uriarte and de Andrés, 2006).

To obtain the highest accuracy, the RF model was optimizedbased on OOB estimate of error rate (Adam et al., 2009; Breiman,2001; Ismail, 2010; Svetnik et al., 2003), using different number oftrees (ntree) from 500 to 10,000 with intervals of 500, while mtrywas optimized using the values between 1 and 20. The ‘‘randomFor-est’’ package (Liaw and Wiener, 2002) developed in R environmentsoftware (R Development Core Team, 2008) was implemented.

Fig. 3. Mean reflectance spectrum data for Hyparrhenia hirta (HH), Eragrostiscurvula (EC), Sporobolus africanus (SA), and Aristida diffusa (AD).

3.2. Forward variable selection

The shortcoming of the random forest algorithm in measuringvariables importance is that it does not automatically select theoptimal number of variables that produce the best classificationaccuracy (Adam et al., 2009). Therefore, forward variable selec-tion (FVS) was used to determine the optimal number of wave-lengths based on the random forest measurement of variablesimportance (Adam et al., 2009; Ismail, 2010). Forward variable

selection iteratively builds multiple random forests (n = 136).At each iteration, two ranked wavelengths were added to themodel, and the error was calculated using the OOB estimatemethod. Initially, the top two wavelengths were selected forthe first iteration, and thereafter the second, third and fourthiteration, the top eight ranked wavelengths were selected. Thisprocess was repeated until no more explanatory variables couldbe included into the final model (Adam et al., 2009).

4. Classification accuracy assessment

To test the prediction performance of any algorithm, the use ofan independent test data set that has not been used in training isrecommended (Congalton and Green, 1999). In random forest algo-rithms, it has been reported that OOB error is considered to be atype of cross-validation that provides an unbiased estimate of error(Archer and Kimes, 2008; Breiman, 2001; Lawrence et al., 2006; Pe-ters et al., 2007). However, some studies have recommended thatthe reliability of OOB estimate of error has to be further tested (Is-mail, 2010; Lawrence et al., 2006). In this study the OOB error wasused to estimate the classification accuracy. Nevertheless, we fur-ther tested the reliability of the OOB error (Lawrence et al., 2006),two methods were used: an independent test data set (n = 22) and

Draw Bootstrap Samples

Input (AISA Eagle (n = 272)) data

Random forest

Trees vote by plurality on the correct classification using the entire variables

Classification accuracy using best variables

R1,R2,R2,R3,R5, R6,R7,R8,R10

R2,R3,R4,R4,R4,R5,R7,R8,R8,R10

R1,R1,R2,R3,R3,R4,R7,R8,R9,R10

R1,R6,R6,R6,R8,R8,R7,R9,R9,R9

Draw Bootstrap Samples

R1,2, R3, R4, R5, R6, R7, R8, …, R272

Input (AISA Eagle (n = 272)) data

Random forest

Tree 3Tree 2 Tree 1 Tree 6500

Measuring variable importance using OOB error

Forward variables selection

Variables ranking using mean decrease in accuracy

Fig. 4. The flowchart describing the random forest model development.

K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65 61

the .632+ bootstrap error for variables selection and classification.The .632+ bootstrap erroris a statistical approach developed by(Efron, 1979) and (Efron and Tibshirani, 1997), and it has beenwidely used for obtaining a nonparametric estimate of error. The.632+ bootstrap error was used with a replication of 50 times toestimate the prediction error at each iteration in forward variablesselection. The optimal number of wavelengths that yielded thesmallest error rate as determined by the three methods (OOB,independent test data set, and the .632+ bootstrap) were then usedto classify the increaser species. A confusion matrix was con-structed to compare the true class with the class assigned by theclassifier and to calculate the overall accuracy as well as the pro-ducer’s and user’s accuracies. The producer’s accuracy is computedby splitting the number of correctly classified trees in each crowncondition class by the number of data sets used for that class(column total in the confusion matrix). User’s accuracy is calcu-lated by dividing the number of correctly classified trees by the to-tal number of trees that were classified in that crown conditionclass (row total in the confusion matrix) (Ismail, 2010).In addition,a discrete multivariate technique was used in accuracy assessment,called Kappa. The result of the Kappa analysis is the KHAT statisticwhich was calculated in order to determine if one error matrix issignificantly different from another (Cohen, 1960). If the Kappa

(K) coefficients are one or close to one then there is perfect agree-ment between the training and test data.

5. Results

5.1. Optimization of ntree and mtry

Following the experiment, the optimization of the number oftrees (ntree) and the number of variables at each split yielded anmtry value of 16 (which is the default setting) and an ntree of6500 resulting in the lowest and stable value of the OOB error rate(14.25%). This optimization result was then used for subsequentanalyses.

5.2. Variables importance using the random forest algorithm

The random forest algorithm effectively explored and describedthe relative importance of each individual wavelength in discrim-inating among increaser species. The most important wavelengthswith the highest mean decrease in accuracy when they are permu-tated are located at 651.9–691.9 nm, 700.8–741 nm and 854.8–966.7 nm (Fig. 5).

Table 3Confusion matrix for eight wavelengths from the training data set showing the classification error obtained for the species (HH, EC, SA and AD). The confusion matrix includesoverall accuracy (ACC), KHAT, commission (CE), user’s accuracy (UA), omission error (OE) and producer’s accuracy (PA).

Species HH EC SA AD Row total CE UA (%)

HH 50 1 1 1 53 5.66 94.34EC 1 48 4 1 54 11.11 88.89SA 1 3 47 2 53 11.32 88.68AD 2 1 2 47 52 9.62 90.38Column total 54 53 54 51 212

OE 7.41 9.43 12.96 7.84 ACC (%) 90.57PA (%) 92.59 90.57 87.04 92.16 KHAT 0.87

Table 4Confusion matrix for eight wavelengths from the test data set showing the classificationerror obtained for the species (HH, EC, SA and AD). The confusion matrix includesoverall accuracy (ACC), KHAT, commission (CE), user’s accuracy (UA) omission error(OE) and producer’s accuracy (PA).

Species HH EC SA AD Row total CE UA (%)

HH 21 1 1 0 23 8.70 91.30EC 0 19 2 1 22 13.64 86.36SA 0 2 18 1 21 14.29 85.71AD 0 1 1 20 22 9.09 90.91Column total 21 23 22 22 88

OE 0 17.39 18.18 9.09 ACC (%) 88.64PA (%) 100 82.61 81.82 90.91 KHAT 0.85

62 K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65

5.3. Forward variable selection

Based on the random forest ranking, the forward variable selec-tion method for the full resampled AISA Eagle wavelengths(nwl = 272) was then used to identify the optimal number of wave-lengths required to discriminate among the species (nsps = 4). Thetop eight wavelengths yielded the lowest OOB error are using thetraining dataset (9.43%), test dataset (11.36%), and .632+ bootstraperror (10.33%) (Fig. 6), compared to the use of the entire wave-lengths (n = 272) that yielded 14.25% (training dataset), 16.05%(test dataset), and 15.95% (.632+ bootstrap error). A t test was usedto test if there was any significant difference among the trainingdata (OOB), test data (OOB) and .632 + bootstrap error. The resultsshow that there is no significant difference between training data(OOB) and test data (OOB) (t = 2.445, P > 0.05) and between train-ing data (OOB) and .632+ bootstrap error. (t = 1.695, P > 0.14).The optimal number of wavelengths that yielded the lowest OOBerror rate (9.43%) and misclassification error based on the .632+bootstrap error measure (15.95%) were selected to classify the spe-cies. These wavelengths (n = 8) are located at 966.7, 877.6, 691.9,718.7, 902.7, 854.8, 674.1 and 703 nm of the electromagnetic spec-trum (the ranking is based on the importance measures).

ig. 5. Identifying the variables (wavelengths) importance by the random forestlgorithm. Wavelengths with the highest mean decrease in accuracy (shown byrrows) represent the most important wavelengths.

5.4. Classification accuracy

Selected optimal wavelengths (n = 8) were used to test the clas-sification accuracy using the confusion matrix derived from theOOB error estimation. The confusion matrix includes overall accu-racy (ACC), KHAT, user’s accuracy (UA), and producer’s accuracy(PA) as shown in Tables 3 and 4. The random forest algorithm usingresampled AISA Eagle data successfully distinguished among in-creaser species (HH, EC, SA, and AD) with an overall accuracy of90.57% and a KHAT value of 0.87 for the training data set(Table 3). An independent test data set was used to test the reli-ability of the OOB error for the classification accuracy with an over-all accuracy of 88.64% and a KHAT value of 0.85 (Table 4). Thedifference in overall accuracy between the training data andindependent test data was less than 3%.

6. Discussion

This study aimed at discriminating increaser species as indica-tors of rangeland degradation using field spectrometry. The moti-vation of the study was to investigate whether there is apossibility to map the different levels of rangeland condition basedon the spatial different distribution of increaser species using theirreflectance spectra. To achieve this, the utility of spectra resampledto AISA Eagle resolution (272 wavelengths) in discriminatingamong four increaser species was tested.

6.1. Optimization of ntree and mtry

Previous studies have shown that RF is sensitive to ntree andmtry parameters (Adam et al., 2009; Díaz-Uriarte and de Andrés,2006; Ismail, 2010; Lawrence et al., 2006). Our result in this studyconfirmed that the high ntree and default setting of mtry(mtry =

pvariables) yielded the lowest of OOB error rate (14.25%).

This result is similar to previous studies (Adam et al., 2009; Brei-man, 2001; Ismail, 2010; Menze et al., 2009; Svetnik et al., 2003)and can be explained by the fact that the highest number of treesallows most of the variables (AISA Eagle bands) to be tested indiscriminating the species (Breiman, 2001).

6.2. Variables importance using the random forest algorithm

The random forest and forward variable selection have success-fully explored and described the relative importance of each indi-vidual wavelength and selected the optimal number ofwavelengths (n = 8) in discriminating increaser species using theOOB method. This optimal number of wavelengths (n = 8) yieldedthe lowest OOB error (9.43%) when compared to the entire wave-lengths (n = 272) which yielded a 14.25% OOB error rate. This canbe explained by the fact that in a model- based analysis, redun-dancy in data can cause convergence instability of models due tonoise in information that has no relation to the increaser speciesbeing classified (Adam et al., 2009; Bajcsy and Groves, 2004; Ismailand Mutanga, 2011). These eight wavelengths are located at 966.7,

Faa

Fig. 6. FVS for the test data set, training data set using OOB, and the bootstrap errorrate. The optimal number of wavelengths that yielded the lowest OOB andbootstrap error is shown by an arrow.

K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65 63

877.6, 691.9, 718.7, 902.7, 854.8, 674.1 and 703 nm and are within±12 nm from known wavelengths that have been reported in pre-vious studies on species discrimination. These are 695, 711 (Chanand Paelinckx, 2008), 689 (Martin et al., 1998), 670 (Daughtryand Walthall, 1998), 675 (Thenkabail et al., 2004), 676, 713, 723(Warner and Shank, 1997), 668, 682, 696, 720 (Thenkabail et al.,2000), 670.37, 695.69, 728.14 (Fung et al., 2003), 692 (Van Aardtand Wynne, 2007) and 720 nm (Vaiphasa et al., 2005).

It can be noted that, all these wavelengths (n = 8) are located inthree different regions of the spectrum which include the visible(n = 1), the red-edge portion (n = 3) and near-infrared portion(n = 4) (Fig. 5). This confirms the results of previous studies thatfound that green leaves have the greatest variation in the visible,red-edge and near-infrared regions (Asner, 1998; Schmidt and Skid-more, 2003; Thenkabail et al., 2004; Vaiphasa et al., 2005). Althoughno leaf biochemical characteristics were directly measured in ourstudy, it is likely that the occurrence of selected wavelengths inthe visible region (400–700 nm) could be due to variation amongthe increaser species on chlorophyll a and b, b-carotene, a-carotene,and xanthophylls (El-Nahry and Hammad, 2009; Ustin et al., 2009).The variations between increaser species in the red-edge region(680–750 nm) due to the chlorophyll concentration, nitrogen con-centration and water content (Ustin et al., 2009). The differencesamong species in the near-infrared region (700–1300 nm) can bethe result of internal leaf structure and water content (El-Nahryand Hammad, 2009; Ustin et al., 2009).

The evaluation of the reliability of the OOB method as an inter-nal estimate of error rate in measuring the importance of thewavelengths using the .632+ bootstrap and test dataset has shownthat this method is reliable. The estimate of error rates from thetest datasets and .632+ bootstrap is nearly identical with a slightdifference of less than 3% to the OOB method. This confirms thefindings in previous studies and supports the assertion that, withrandom forest, it is not necessary to have an independent test dataset (Lawrence et al., 2006).

6.3. Classification accuracy

The optimal wavelengths (n = 8) yielded an overall accuracy of90.57% and a KHAT value of 0.87 and the producer’s accuracy rangedfrom 87.04% to 92.59% and the user’s accuracy ranged from 88.68%to 94.34% (Table 3). The results offer the possibility of classifyingand mapping rangeland degradation with high classification accu-racy (90.57%) based on the distribution of the increaser species.

The reliability of the OOB error for the classification accuracywas tested using an independent test data set which yielded anoverall accuracy of 88.64% and a KHAT value of 0.83. The differencein overall accuracy between the training data set and the indepen-dent test data was less than 3% (Table 4) which confirm the stabil-ity and reliability of the OOB error (Lawrence et al., 2006). The useof the internal error measure could save field data collection timeby reducing the number of the samples to be collected for validat-ing the performance of RF (Lawrence et al., 2006).

In summary, the results presented in this study confirm that theRF algorithm is a robust and accurate method for the combinedpurposes of variables selection and the classification of hyperspec-tral data. Overall, this study demonstrated the possibility of dis-criminating increaser species using resampled data. This allowsthe upscaling of methods to airborne sensors such as AISA Eaglefor mapping rangeland degradation using increaser species asindicators.

7. Conclusions

This paper aimed at discriminating among four increaser spe-cies: Hyparrhenia hirta,Eragrostis curvula, Sporobolus africanus andAristida diffusa using field spectrometry data, resampled to AISAEagle resolution. Our results have shown that:

(1) These increaser species have a strong potential to be classi-fied accurately using spectrometry data.

(2) The random forest algorithm has several advantages, such asbeing able to provide better performance, reasonable accu-racies, and ease of use.

(3) The random forest algorithm, using hyperspectral data, dis-criminated among four increaser species with a high accu-racy of 88.64% (KHAT 0.85).

(4) The random forest algorithm has revealed that greater dis-crimination power is contained in the visible, red-edge andnear-infrared regions of the spectrum. The optimal numberof wavelengths that yielded the lowest OOB error rate areat 966.7, 877.6, 691.9, 718.7, 902.7, 854.8, 674.1 and 703 nm.

The results demonstrated the possibility of discriminatingincreaser species using hyperspectral data, resampled to anairborne sensor. This permits the possibility to upscale the meth-ods to airborne sensors such as AISA Eagle for mapping increaserspecies areas as an indicator of rangeland condition.

Acknowledgements

Thanks are due to Brice Gijsbertsen a chief Cartographer(Department of Geography, University of KwaZulu-Natal) for assis-tance in configuring the ASD sensor. My gratitude goes to AbdallahIbrahim and Dasali for their support during the data collection andto Susan Davies for editing.

References

Adam, E., Mutanga, O., 2009. Spectral discrimination of papyrus vegetation (Cyperuspapyrus L.) in swamp wetlands using field spectrometry. ISPRS Journal ofPhotogrammetry and Remote Sensing 64 (6), 612–620.

Adam, E.M., Mutanga, O., Rugege, D., Ismail, R., 2009. Field spectrometry of papyrusvegetation (Cyperus papyrus L.) in swamp wetlands of St Lucia, South Africa. In:Geoscience and Remote Sensing Symposium, 2009 IEEE International, IGARSS2009, pp. IV-260-IV-263.

Adger, W.N., Benjaminsen, T.A., Brown, K., Svarstad, H., 2000. Managing FragileEcosystems: Combating Desertification and Drought. Centre of Social andEconomic Research on the Global Environment, University of East Anglia,London.

Archer, K., Kimes, R., 2008. Empirical characterization of random forest variableimportance measures. Computational Statistics and Data Analysis 52 (4), 2249–2260.

64 K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65

ASD, Analytical Spectral Devices, I., 2005. Analytical Spectral Devices, Inc., HandheldSpectroradiometer: User’s Guide, Version 4.05, Boulder, USA.

Asner, G., 1998. Biophysical and biochemical sources of variability in canopyreflectance. Remote Sensing of Environment 64 (3), 234–253.

Bajcsy, P., Groves, P., 2004. Methodology for hyperspectral band selection.Photogrammetric Engineering and Remote Sensing 70 (7), 793–802.

Bangamwabo, V.M., 2009. Spatial and temporal extents of land degradation in acommunal landscape of KwaZulu-Natal. University of KwaZulu-Natal,Pietermaritzburg Geography.

Breiman, L., 2001. Random forests. Machine learning 45 (1), 5–32.Chan, J., Paelinckx, D., 2008. Evaluation of random forest and adaboost tree-based

ensemble classification and spectral band selection for ecotope mapping usingairborne hyperspectral imagery. Remote Sensing of Environment 112 (6), 2999–3011.

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

Cohen, J., 1960. A coefficient of agreement for nominal scales. Educational andpsychological measurement 20 (1), 37–46.

Congalton, R., Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data:Principles and Practices. Lewis, Boca Raton, FL.

Critchley, W., Netshikovhela, E., 1998. Land degradation in South Africa:conventional views, changing paradigms and a tradition of soil conservation.Development Southern Africa 15 (3), 449–470.

Daughtry, C., Walthall, C., 1998. Spectral discrimination of Cannabis sativa L. leavesand canopies. Remote Sensing of Environment 64 (2), 192–201.

Díaz-Uriarte, R., de Andrés, A., 2006. Gene selection and classification of microarraydata using random forest. BMC bioinformatics 7 (1), 3.

Dollar, E., Goudy, A., 1999. Environmental change. The geography of South Africa ina changing world. Oxford University Press, Oxford, 477p.

du Toit, J.C.O., 2009. Early survival and growth of vegetatively propagatedindigenous grasses in a clear-felled timber plantation in KwaZulu-Natal,South Africa. African Journal of Range and Forage Science 26 (2), 97–101.

Efron, B., 1979. Bootstrap methods: another look at the jackknife. The annals ofStatistics 7 (1), 1–26.

Efron, B., Tibshirani, R., 1997. Improvements on cross-validation: the. 632+bootstrap method. Journal of the American Statistical Association 92 (438),548–560.

El-Nahry, A., Hammad, A., 2009. Assessment of salinity effects and vegetation stress,West of Suez Canal, Egypt using remote sensing techniques. Journal of AppliedSciences Research 5 (3), 316–322.

Everson, C., Tainton, N., 2007. The effect of thirty years of burning on the HighlandSourveld of Natal. African Journal of Range and Forage Science 1 (1), 15–20.

Fung, T., Yan, H., Siu, W., 2003. Band selection using hyperspectral data ofsubtropical tree species. Geocarto International 18 (4), 3–11.

Hill, J., Stellmes, M., Udelhoven, T., Röder, A., Sommer, S., 2008. Mediterraneandesertification and land degradation: Mapping related land use changesyndromes based on satellite observations. Global and Planetary Change 64(3–4), 146–157.

Hoffman, M., Todd, S., 2000. A national review of land degradation in South Africa:the influence of biophysical and socio-economic factors. Journal of SouthernAfrican Studies 26 (4), 743–758.

Hoffman, M., Todd, S., Ntshona, Z., Turner, S., 1999. Land degradation in South AfricaDepartment of Environment Affairs and Tourism. Pretoria.

Ismail, R., 2010. Remote sensing of forest health: the detection and mapping ofPinus patula trees infested by Sirex noctilio. Ph.D. Thesis, School ofEnvironmental Sciences, University of KwaZulu-Natal, Pietermaritzburg, SouthAfrica.

Ismail, R., Mutanga, O., 2011. Discriminating the early stages of Sirex noctilioinfestation using classification tree ensembles and shortwave infrared bands.International Journal of Remote Sensing 32 (15), 4249–4266.

Lawrence, R., Wood, S., Sheley, R., 2006. Mapping invasive plants usinghyperspectral imagery and Breiman Cutler classifications (RandomForest).Remote Sensing of Environment 100 (3), 356–362.

Liaw, A., Wiener, M., 2002. Classification and regression by random forest. R News 2(3), 18–22.

Lorent, H., Evangelou, C., Stellmes, M., Hill, J., Papanastasis, V., Tsiourlis, G., Roeder,A., Lambin, E., 2008. Land degradation and economic conditions of agriculturalhouseholds in a marginal region of northern Greece. Global and PlanetaryChange 64 (3–4), 198–209.

Martin, M., Newman, S., Aber, J., Congalton, R., 1998. Determining forest speciescomposition using high spectral resolution remote sensing data. RemoteSensing of Environment 65 (3), 249–254.

Menze, B., Kelm, B., Masuch, R., Himmelreich, U., Bachert, P., Petrich, W., Hamprecht,F., 2009. A comparison of random forest and its Gini importance with standardchemometric methods for the feature selection and classification of spectraldata. BMC bioinformatics 10 (1), 213.

Muchoney, D., Haack, B., 1994. Change detection for monitoring forest defoliation.Photogrammetric Engineering and Remote Sensing 60 (10), 1243–1251.

Mutanga, O., 2005. Discriminating tropical grass canopies grown under differentnitrogen treatments using spectra resampled to HYMAP. International Journalof Geoinformatics 1 (2), 21–32.

Mutanga, O., Skidmore, A., 2004. Hyperspectral band depth analysis for a betterestimation of grass biomass (Cenchrus ciliaris) measured under controlledlaboratory conditions. International Journal of Applied Earth Observation andGeoinformation 5 (2), 87–96.

Mutanga, O., Skidmore, A., Kumar, L., Ferwerda, J., 2005. Estimating tropical pasturequality at canopy level using band depth analysis with continuum removal inthe visible domain. International Journal of Remote Sensing 26 (6), 1093–1108.

Nordberg, M., Allard, A., 2002. A remote sensing methodology for monitoring lichencover. Canadian journal of remote sensing 28 (2), 262–274.

O’Connor, T.G., Bredenkamp, G.J., 1997. Grassland. In: Cowling, R.M., Richardson,D.M., Pierce, S.M. (Eds.), Vegetation of Southern Africa. Cambridge UniversityPress, Cambridge, pp. 215–257.

Oluwole, F., Sambo, J., Sikhalazo, D., 2008. Long-term effects of different burningfrequencies on the dry savannah grassland in South Africa. African Journal ofAgricultural Research 3 (2), 147–153.

Özçift, A., 2011. Random forests ensemble classifier trained with data resamplingstrategy to improve cardiac arrhythmia diagnosis. Computers in Biology andMedicine 41 (5), 265–271.

Palmer, A., van Rooyen, A., 1998. Detecting vegetation change in the southernKalahari using Landsat TM data. Journal of Arid Environments 39 (2), 143–153.

Peters, J., De Baets, B., Samson, R., Verhoest, N., 2007. Modelling groundwater-dependent vegetation patterns using ensemble learning. Hydrology and EarthSystem Sciences Discussions 4 (5), 3687–3717.

Prince, S., Becker-Reshef, I., Rishmawi, K., 2009. Detection and mapping of long-term land degradation using local net production scaling: application toZimbabwe. Remote Sensing of Environment 113 (5), 1046–1057.

Pyke, D., Herrick, J., Shaver, P., Pellant, M., 2002. Rangeland health attributes andindicators for qualitative assessment. Journal of Range Management 55 (6),584–597.

Ravi, S., Breshears, D., Huxman, T., D’Odorico, P., 2010. Land degradation indrylands: interactions among hydrologic-aeolian erosion and vegetationdynamics. Geomorphology 116 (3–4), 236–245.

R Development Core Team, 2008. R: a language and environment for statisticalcomputing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, available online at: <http://www.R-project.org> (accessed21.01.2011).

Reid, P., Vogel, C., 2006. Living and responding to multiple stressors in South Africa-Glimpses from KwaZulu-Natal. Global Environmental Change 16 (2), 195–206.

Schmidt, K., Skidmore, A., 2003. Spectral discrimination of vegetation types in acoastal wetland. Remote Sensing of Environment 85 (1), 92–108.

Shackleton, C., Guthrie, G., Main, R., 2005. Estimating the potential role ofcommercial over harvesting in resource viability: a case study of five usefultree species in South Africa. Land Degradation and Development 16 (3), 273–286.

Svetnik, V., Liaw, A., Tong, C., Culberson, J., Sheridan, R., Feuston, B., 2003. Randomforest: a classification and regression tool for compound classification and QSARmodeling. Journal of Chemical Information and Computing Science 43 (6),1947–1958.

Temme, A., 2008. Understanding Landscape Dynamics over Thousand Years:Combining Field and Model Work: With Case Study in the DrakensbergFoothill. KwaZulu-Natal, South Africa.

Thenkabail, P., Enclona, E., Ashton, M., Van Der Meer, B., 2004. Accuracyassessments of hyperspectral waveband performance for vegetation analysisapplications. Remote Sensing of Environment 91 (34), 354–376.

Thenkabail, P., Smith, R., De Pauw, E., 2000. Hyperspectral vegetation indices andtheir relationships with agricultural crop characteristics. Remote Sensing ofEnvironment 71 (2), 158–182.

Thompson, M., 1996. Standard land-cover classification scheme for remote-sensingapplications in South Africa. South African Journal of Science 92 (1), 34–42.

Trollope, W., 1990. Development of a technique for assessing veld condition in theKruger National Park using key grass species. Journal of the Grassland Society ofSouthern Africa 7 (1), 46–51.

UNCCD, 1995. United Nations Convention to Combat Desertification in CountriesExperiencing Serious Drought and/or Desertification, Particularly in Africa, textwith Annexes. UNEP, Geneva.

Ustin, S., Jacquemoud, S., Palacios-Orueta, A., Li, L., Whiting, M., 2009. Remotesensing based assessment of biophysical indicators for land degradation anddesertification. Recent advances in remote sensing and geoinformationprocessing for land degradation, assessment, 15.

Vaiphasa, C., Ongsomwang, S., Vaiphasa, T., Skidmore, A., 2005. Tropical mangrovespecies discrimination using hyperspectral data: a laboratory study. Estuarine,Coastal and Shelf Science 65 (1–2), 371–379.

Vaiphasa, C., Skidmore, A., de Boer, W., Vaiphasa, T., 2007. A hyperspectral bandselector for plant species discrimination. ISPRS Journal of Photogrammetry andRemote Sensing 62 (3), 225–235.

Van Aardt, J., Wynne, R., 2007. Examining pine spectral separability usinghyperspectral data from an airborne sensor: an extension of field-basedresults. International Journal of Remote Sensing 28 (1–2), 431–436.

Van Lynden, G., Mantel, S., 2001. The role of GIS and remote sensing in landdegradation assessment and conservation mapping: some user experiences andexpectations. International Journal of Applied Earth Observation andGeoinformation 3 (1), 61–68.

Van oudtshoorn, F.P., 1992. Guide to Grasses of South Africa. Briza, Pretoria.Vesk, P.A., Westoby, M., 2001. Predicting plant species’ responses to grazing. Journal

of Applied Ecology 38 (5), 897–909.Wang, J., He, T., Lv, C., Chen, Y., Jian, W., 2010. Mapping soil organic matter based on

land degradation spectral response units using Hyperion images. InternationalJournal of Applied Earth Observation and Geoinformation 12 (2), S171–S180.

Warner, T., Shank, M., 1997. Spatial autocorrelation analysis of hyperspectralimagery for feature selection⁄ 1. Remote Sensing of Environment 60 (1), 58–70.

K. Mansour et al. / ISPRS Journal of Photogrammetry and Remote Sensing 70 (2012) 56–65 65

Wessels, K., Prince, S., Frost, P., Van Zyl, D., 2004. Assessing the effects of human-induced land degradation in the former homelands of northern South Africawith a 1 km AVHRR NDVI time-series. Remote Sensing of Environment 91 (1),47–67.

Wessels, K.J., Prince, S., Malherbe, J., Small, J., Frost, P., VanZyl, D., 2007. Can human-induced land degradation be distinguished from the effects of rainfall

variability? A case study in South Africa. Journal of Arid Environments 68 (2),271–297.

Wessels, K.J., Prince, S.D., Reshef, I., 2008. Mapping land degradation by comparisonof vegetation production to spatially derived estimates of potential production.Journal of Arid Environments 72 (10), 1940–1949.