Published on

International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.

Published in: Technology
1 Like
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693686 | P a g eFeature Extraction and Based Pixel Classification for Estimationthe Land Cover thematic map using Hyperspectral data*Hayder Abd Al-Razzaq Abd*(Department of Remote Sensing and GIS unit, Civil Engineering, UPM University, Malaysia)ABSTRACTThere are a few studies have been usedthe hyperspectral remotely sensed imageries toidentify the Land use and land cover into UPMUniversity campus. UPM is one of the oldestMalaysian Universities located in the westMalaysia, Selangor state, Serdang it chosen forthis research. High resolution images fromhyperspectral satellite and Google earth havebeen used to conduct this research. The studyarea divided into nine classes (Clear water,Lake, Soil, Roads, Building R- roof, Building C-roof, Building B- roof, Grass, Tree) to estimatethe thematic map of land cover of the UPMcampus, using two algorithms the classificationperformed with Support Vector Machine underfeature extraction classification and Maximumlikelihood under based pixel classification, thendid the comparison between the two techniques.Through the whole study duration. The resultsdemonstrated that the accuracy of SVM is betterthan of Maximum likelihood and the overallaccuracy for SVM and Maximum likelihoodwere 98.23%, 90.48% respectively.Keywords – hyper spectral image, support vectormachine, Maximum likelihood, feature extraction,based pixel classification. land cover.1. INTRODUCTIONLand cover and land use is the discernibleto the Vegetation, Geologic, Anthropogenic orHydrologic features on the earths land surface.Land cover describes the physical state of theearth’s surface and immediate surface in terms ofthe natural environment and the manmadestructures (such as water bodies, vegetation, urbanarea, soils, earth’s surfaces and ground water).These land cover features can be classified usingthe remotely sensed satellite imagery of differentspectral, temporal and spatial resolutions. Landcover mapping using high spectral resolution hasmany advantages, since it helps in different kind ofmapping applications such as speciesidentifications, soil types, mineral classificationsetc. The hyperspectral data has a large number ofcontiguous bands that have high level of spectralresolution. The hyperspectral data processingpossesses both challenges and opportunities forland cover classification. Land cover classificationcan be conducted using various algorithms byprocessing the remotely sensed data into differentthemes. The present research examines andcompares the results obtained from variousclassification algorithms using hyperspectral datafor land cover mapping to find out which algorithmand technique better for such a land cover mapping.2. BACKGROUNDHyperspectral remote sensing (HRS)images can be acquired from different kind ofhyperspectral sensors, such as AVIRIS, EO-1Hyperion, ASIA and HyMap, have revealed verywide usefulness in many applications (environment, agriculture and geosciences).Generally, the hyperspectral remotely sensingimage has massive information and the fine spectralresolution which can capture the subtle differencesbetween the spectral reflectance of differentvegetation species [1], that will make thehyperspectral remote sensing data processing isvery complex in the terms of data uncertainty [2],curse of dimensionality [1] and small samples [3,4],as a result of that it is very difficult to employtraditional or normal type of classification withhyperspectral data with getting as output highperformance. The Huge of any phenomenon thatdetected with hayperspectral imagery leads toserious problem related to decrease the accuracy ofthe classification for this phenomenon to extract theinformation [5]. Support vector machine (SVM)has been shown to outperform classical supervisedclassification algorithms, therefore it has beenrecently used for classification of hyperspectraldata. In comparison with approaches based onempirical risk, which minimize themisclassification error on the training set, structuralrisk minimization seeks the smallest probability ofmisclassifying a previously unseen data pointdrawn randomly from a fixed but unknownprobability distribution. Furthermore, an SVM triesto find an optimal hyperplane that maximizes themargin between classes by using a small number oftraining cases. The basis of the SVM and the resultsof some studies suggest that SVM classificationmay be unaffected by the dimensionality of the dataset, however, some studies have shown that theaccuracy of SVM classification could still beincreased by reducing the dimensionality of thedata set. The SVM has been widely used andpromoted for land-cover classification studies,
  2. 2. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693687 | P a g eincluding multispectral and hyperspectral data withsome studies suggesting that the method is notaffected by the Hughes phenomena. However, arecent research shows that the accuracy of SVMclassification is influenced by the number offeatures used, therefore, is affected by the Hughesphenomenon with the impact most evident when asmall training set is used. It is possible that theHughes effect had not been observed in some otherstudies because the opportunity for it to becomemanifested in the results was limited throughexperimental design, notably through the use of alarge training set. Two broad categories of feature-reduction techniques: they are feature extractionand feature selection, with feature extraction, theoriginal remotely sensed data set is typicallytransformed in some way that allows the definitionof a small set of new features which contain thevast majority of the original data set’s information.More popular, are feature-selection methods. Thelatter aim to define a subset of the original featureswhich allows the classes to be discriminatedaccurately. Feature selection typically aims toidentify a subset of the original features thatmaintains the useful information to separate theclasses with highly correlated and redundantfeatures excluded from the classification analysis.The results of the same research have shown thatthe accuracy of classification by an SVM can besignificantly reduced by the addition of featuresand that the effect is most apparent with smalltraining sets. With the AVIRIS data set, asignificant reduction in accuracy with the additionof features was observed at all training set sizesevaluated. With the DAIS data set, a statisticallysignificant decline in accuracy was also observedfor small training sets (≤ 25 cases per class).However, even with a large training sample usingthe DAIS data set, feature selection may have apositive role, providing a reduced data set that maybe used to yield a classification of similar accuracyto that derived from use of much larger feature set.As the accuracy of SVM classification wasdependent on the dimensionality of the data set andthe size of the training set, it may be beneficial toundertake a feature-selection analysis prior to aclassification analysis. Another study compared theresults of SVM and maximum likelihood classifier(MLC) in classification to provide a land use/covermap in an urban area. A broadband image wassimulated and derived a band reduced image fromthe AVIRIS image. The simulated broadbandimage provides the identical spectral coverage ofthe six TM bands (excluding the thermal band).The band reduced image was created by using datadimensionality reduction technique to select 85AVIRIS bands, which allow performing someclassification types (e.g., maximum likelihoodclassifier) to be effectively performed withouthaving to substantially increase the amount oftraining data. The hyperspectral imageries contain ahuge amount of redundant spectral information inthe adjacent bands, and the principal componentanalysis (PCA) method was used to reduce the datadimension, and the output image consists of sixPCA bands that contain 99.9% of the informationfrom the 85 AVIRIS bands. A land use-coverclassification scheme based on the Andersonscheme [21] was designed, which includes 7 majorcategories: low-density urban, medium-densityurban, high-density urban, forest, grassland, barrenland and water. The three urban land classes varyby their built-up densities. Training and referencesamples were selected through the use of high-resolution images from Google Earth and NationalLand Cover Data.The result summarizes the accuracyassessment for each image classified with SVM andmaximum likelihood methods. The AVIRIS imagewith 85 bands has the higher overall classificationaccuracy when classifying with the SVM method.Note that the researcher failed in classifying thesame image with the maximum likelihood (MLC)method, largely due to the inability of thisconventional method in dealing with the substantialdata dimensionality increase with relatively limitedtraining sample size. Support vector machines(SVM) substantially outperformed the maximumlikelihood classifier in each image. In terms ofspecific land cover classes, the AVIRIS data whichclassified using the SVM method shown a betteraccuracy, particularly for the three urban classes,when comparing with the outcomes from otherimages, the SVMs generally outperformed theMLC method. Overall, the improved capability ofhyperspectral imagery helped better resolvespectrally complex classes, particularly whenclassified with the support vector machine method.Support Vector Machine is one of the bestclassifiers that adopted to perform the classificationof hyperspectral remotely sensed imageriesclassification successfully. The result or the outputof the Support Vector Machine show that thisclassifier is better than another classifiers such asMinimum Distance Classification (MDC),Maximum likelihood Classifications (MLC)Artificial Neural Network (ANN), and SpectralAngle Mapper Classification (SAM) [6, 7, 11].Recently SVM became the most popularand effective statistical learning algorithm that usein pattern recognition and machine learning fields,SVM has such advantages as less requirement toprior knowledge, more suitability to small size ofsamples [8], more robustness to noises [9], andhigher learning efficiency and more powerfulgeneralization capacity [10]. So it can be used tospatial data processing and analysis includinghyperspectral RS image classification [11], spatialfitting and regression [12], data mining [13], andobject detection [14]. SVM, as one of the most
  3. 3. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693688 | P a g eeffective statistical algorithms, it is considered thestructural risk minimization (SRM) criterion morethan considered empirical risk minimization (ERM)in other methods. Because SVM is advantageous toreduce difficulties that will face the hyperspectraldata when classifying a small-size samples, highdimensionality, poor generalization and uncertaintyimpacts, it has been applied to hyperspectralRemote Sensing image classification in recent years[15, 16]. The SVM is perfect classifier algorithmthan other traditional classifiers and it is perfect forhigh dimensional features such as the direction usefor hyperspectral image’s bands but the wastingtime and the computation capacity are still bigchallenges, therefor, feature extraction anddimensionality decreasing still meaningful on manycase.The Maximum Likelihood classifierassumes that the statistics for each class in eachband are normally distributed and calculates theprobability that a given pixel belongs to a specificclass. Unless a probability threshold is selected, allpixels are classified, each pixel is assigned to theclass that has the highest probability. If the highestprobability is smaller than a threshold, the pixelremains unclassified. [17] In the study which wasdone in 2007, different classifications such asArtificial Neural Network (ANN) classifier,Spectral Angle Mapper (SAM), Decision Tree (DT)classifier, Spectral Angle Mapper (SAM) andMaximum Likelihood Classifier (MLC) wereimplemented in a hyperspectral image in Malaysia.The higher accuracy of results as shown by theMLC classifier, this study suggests that thehyperspectral data which was derived from theoptimal band configuration of the airborne sensorhas a sufficiently Gaussian distribution that is ableto give a full and representative description of therespective classes (spectrally separable tree speciesclasses) and fulfills the requirement (biasedtowards) for such a parametric algorithm [17]. Theproblem of not having enough training sampleswith hyperspectral imagery for reliable training of amaximum likelihood classifier is well known.Another research shows that the providedcandidate neighbors which are assessed forsuitability spectrally can be used to expand thetraining set and generate class statistics for highdimensionality data giving improvements inclassifier performance comparable to that observedwith randomly selected unlabeled samples, alongwith a more complex evaluation process [18]. Thestudy which was done in 2009 it found the maindifficulty of texture recognition is the lack ofeffective tools to characterize different scales oftextures and then to improve the problem, thewavelet co-occurrence parameters, mean,homogeneity, and standard deviation of differentlevel discrete wavelet transform images were usedas texture features, then the texture featurescombined with PCA band of image were adopted asthe characteristic vector of training samples forSVM, and Decision Tree and Maximum likelihoodclassification. The experimental results showed thatSVM method gave the highest correct classificationrate within all of these three methodologies whileMaximum Likelihood gave the lowest rate and alsoadding texture feature information by the proposedapproach to images improved classificationaccuracy for all of SVM, Decision Tree, andMaximum Likelihood classification [19].3. METHODS AND MATERIALS3.1. Study AreaThe study area was in campus ofUniversity Putra Malaysia (UPM) that located inSerdang, Selangor state, Malaysia. UPM universityone of the biggest and oldest universities inMalaysia, it has been established in 1931, UPM is aresearch university in central Peninsular Malaysia,near to the biggest and the capital city, KualaLumpur. It was formerly known as UniversityPertanian Malaysia or Agricultural University ofMalaysia or University Putra Malaysia. UPM is avery good research university providingundergraduate and postgraduate courses with aresearch focus on different sciences. The Fig. 1below describes the study area.Fig. 1: The study area in UPM University Campus.3.2. DataThe data has been used to conduct thisstudy was hyperspectral remotely sensed imageryhas 20 bands this imagery is represented the studyarea in UPM campus .The Area Of study area was(729,828) m2.The hyperspectral imagery has beengeo-referenced to coordinate system {UTM, Zone47N, with spatial resolution 1 meter and Fig. 2reveals the hyperspectral image
  4. 4. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693689 | P a g eFig. 2: Hyperspectral imagery of study areaThe datum was (WGS-84) .The wavelength ofthe 20 bands between Wavelength between{443.7250, 889.1100} in other words is include thevisible and near infrared of electromagneticspectrum and the Table below describes thelocation of study area with using WGS 84 as aDatum. The date of capturing this imagery was inMay, 2004. The method that used to perform thisstudy reveals in the flow chart in Fig. 3.Table 1: The coordinate system of the study areaboundaries with WGS-84 Datum.Fig. 3: Flow chart of the methodology.3.3. Image analysis3.3.1. Subset the digital imageryThe method that the researcher followedfor this study it was after the investigated andinterpreted the hyperspectral imagery he found thatthere was some overlapping with the boundaries ofthe image, that was important issue should dealwith to overcome this problem the researcher sub-setted the image to remove this noise from theimagery by using ENVI software and by selectedthe area that out of the noise using the spatial subsetoption then the image has been subsetted. The sub-setting the hyperspectral imagery done before doingany analysis.3.3.2. Atmospheric correction:In this stage because the hyperspectral imagerydid not correct, and the digital imagery is related toold date, in another words because there is not anyfield parameters to do the advance atmosphericcorrection, the internal Average RelativeReflectance (IARR) has been applied to performingthe atmospheric correction. This step also did withusing the Envi software.3.3.3. Training and Testing Sites Selection.The training sites collected from imagery byselecting the region of interest (ROIs) using Envisoftware, the study area classified into ninecategories, therefore, nine ROIs were collectedfrom the hyperspectral imagery and the testing sitescollecting using Global Positioning System (GPS)and Google earth from the historical database ofGoogle earth (archive), testing sites selected and itrepresented the truth samples that represent theseclasses. Selection testing sites very important to dothe assessment of classification result, then studyarea was classified into (9) classes using twoapproaches the first one is normal supervisedclassification (per – pixel classification) it wasMaximum likelihood classifier (MLC) to conductthe classification of study area, and also applied theSupport Vector Machine (SVM) based on thespatial and the spectral aspect (object orientedclassification) both of them have been used togenerate the thematic map of land cover for UPMcampus and then performed the generalization toboth of outputs classification to isolate theunclassified pixel and to make the map has a goodvisualization . By using the testing sites thatcollected from Google earth the accuracyassessment of classification result for bothclassifiers done then the final step it was make acomparison between the results of both classifiersto demonstrates the better technique that can beused to extract better result for land cover mapping.4. RESULT AND DISCUSSION4.1 Image Analysis Processing4.1.1 Subset the digital imageryFirst step in this Study was subsetted thehyperspectral imagery to remove and overcome thenoise that appeared on the boundaries and thesubsetted was conducted with using ENVIsoftware. The result of this subsetted wasdemonstrates in Fig. 4.
  5. 5. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693690 | P a g eFig.4: The sub-setted of hyperspectral imagery.4.1.2 Atmospheric correction:After the sub-setting the hyperspectral imagerythe next step was performed the atmosphericcorrection, the technique that has been used wasinternal Average Relative Reflectance (IARR) andthat because the hyperspectral image related to2004 and it is old and there is no any filedmeasurements to use as variable input parametersto do high atmospheric correction (field spectralmeasurements and other observations) related tostudy area in time of over flight the researcherselected this IARR technique for performing theAtmospheric correction and the Fig. 5 below showsthe spectral reflectance before and after conductedthe atmospheric correction.Fig. 5: Spectral reflectance before and afteratmospheric correction.4.1.3. Training Sites SelectionSelecting training is the next step in theanalysis, the training sites collect to each type ofnine classes of study area by collecting the regionof interests (ROIs) for each class to use thesetraining sites to perform the classification then toobtain the thematic map of Land cover, theseclasses are: (Grass, Trees, Clear water, Lake,Roads, Building R-roof, Building C-roof andBuilding B-roof) we have three classes of buildingsbased on their roof’s material. The Table 2 revealsthe training sites of the study area.Table 2: Training sites of study area.4.1.4. Testing Sites SelectionThe testing sites are very important to conductthe analysis especially to do the accuracyassessment for both of classifications (SVM andMaximum likelihood), with using the high spatialresolution IKONAN from Google Earth the studyarea was investigated to select testing sites for thefeatures (Soil, buildings roofs and lakes) to use it inaccuracy assessment stage, the samples collectedby selected some from the historical archive ofGoogle earth and by field work with using theGlobal Positioning System (GPS) hand held GPSthere were some ground truth samples collected forfeatures (soil, swimming pool, Roads, trees andGrass), GPS has the accuracy (-,+5) meters, therewere 18 ground samples were collected inside thestudy area in UPM university campus as indicatedin Table 3, these ground truth samples werecollected based on the datum WGS84 and Fig. 6indicates the positions of these samples.Fig 6: Samples of testing sites in UPM Campus.
  6. 6. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693691 | P a g eTable 3: Demonstrates the Ground truth samples inUPM campus that collected by GPS.4.1.5. ClassificationThis research conducting with two typesof classification the first one is Maximumlikelihood algorithm based per pixel and using theROIs that collected before to generate the thematicmap of Land Cover of UPM Campus for nineclasses. The second was SVM based on objectoriented classification, the segmentation of imagedone then by selecting the kernel type that has beenused was Radial base function type for conductingthe classification.4.1.5 Post classification (Sieving and Clumping)The post classification is an importantstage to enhance the quality of classificationsresults. It plays a big role include the recoding ofland cover classes, and some modification of theclassified remotely sensed imagery by usinganother information such as ancillary data or expertknowledge. Most of classification that done forland cover have some salt and pepper effectsrelated to the complexity of landscape, that needsto apply some filters to remove that effect, and thisprocedure can be done by using low pass filter ordo some another kind of smoothing like Clumpingand Sieving techniques. The parameters had beenchosen for SVM and MLC, classificationsperformed, then the generalization the classifiedimage done using the post classification withSieving and clumping to remove the isolatedpixels and to give the thematic map bettervisualization and the result for both classificationswere as the Fig. 7,8 below:After generalization before generalizationFig. 7: Thematic map of SVM classification.Fig 8: The generalization of Maximum likely hoodclassification.And here below Fig. 9 indicates the finalthematic map of land use and land cover for UPMCampus using Maximum likelihood algorithm thathas been generated by using Envi software.Fig 9: The final thematic map LU_LC for UPMCampus by MCL.Fig 10: The final thematic map of LU_LC for UPMCampus by SVM.
  7. 7. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693692 | P a g e4.1.6. Accuracy assessmentIn this stage the accuracy assessment fortwo techniques calculated by confusion matrixerror and by using the ground truth sites that havebeen collected from historical archive of Googleearth to perform the assessment. The Accuracyassessment use to determine the quality ofinformation generated from remotely sensed data.There are several reasons for conducting theassessment of accuracy to remotely sensed data.They are:1) To know how much good the results.2) To determine and then to correct sources oferror.3) To make comparison between differenttechniques, algorithms, analysts or interpretersthat have been used.4) Check the quality of the results.After performing the accuracy assessmentthe results were the Overall Accuracy = 90.48%and Kappa Coefficient = 0.88 for Maximum likelyhood classification however, for SVMclassification the results were Overall Accuracy =98.23% and Kappa Coefficient = 0.97, and thecommission, omission percentage, produceraccuracy and user accuracy demonstrated as thetables (4), (5), (6), (7) below:Table 4: The percentage of commission andomission errors of MLC.Table 5: The percentage of Prod.Acc. & User Acc.of MLC.Table 6: The percentage of commission andomission errors of SVM.Table 7: The percentage of Prod.Acc. & User Acc.of SVM.4.1.7. ComparisonThe comparison result was indicated that theSVM classification is better than the MLC and theTable 8 and Fig. 11 below indicate that:Table 8: The Comparison results between SVM &MCL.The different in the accuracy assessmentrevealed that the object oriented classification isbetter than the normal classification (per pixel),that related to the classifier with the objectedoriented approach will classify the imagery basedon the spatial, spectral and the texturecharacteristics of the features that will support theprocess of the classification and increase theparameters of classification then the output will bebetter identification than the per pixel approach,because it considers just the spectral characteristicsof feature.
  8. 8. Hayder Abd Al-Razzaq Abd / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.686-693693 | P a g eFig 11: comparison chart between MCL and SVMaccuracy results.5. CONCLUSIONThis study covers the processing ofhyperspectra data to classify the study area thatlocated in UPM University in Serdang, Selangor,Malaysia with using two methods the first oneswith SVM algorithm under feature extraction andthe second using MCL under the classificationbased on pixel to find out the thematic map of landcover and examine which type of classification isbetter. And the results revealed that the accuracyassessment of SVM is better than of MLCclassification and the overall accuracy for both ofthem, for SVM and Maximum likelihood were98.23%, 90.48% respectively. The high results thatgot reflect the ability of these algorithms to dealwith the land cover and land use mapping even thearea is very small and there is a good variation inthe type of features that distributed in study areathat will help to get good accuracy and make theidentifications for different classy more easier.6. ACKNOWLEDGEMENTSThe author would like to thank the Assoc. Prof.Dr. Helmi for his support and providing the data,and also the author very grateful from Mr. Eng.Maher Faiq Ismaeel for his supporting.REFERENCES[1] F. Melgani, L. Bruzzone, IEEETransactions on Geoscience and RemoteSensing 42, 2004,pp 1778.[2] T. Oommen, D. Misra, N. Twarakavi, A.Prakash, B. Sahoo, S. Bandopadhyay,Mathematical Geosciences 40, 2008,pp409.[3] G. Foody, A. Mathur, Remote Sensing ofEnvironment 93, 2004,pp 107.[4] G. Foody, A. Mathur, C. Sanchez-Hernandez, D. Boyd, Remote Sensing ofEnvironment 1042006, 1.[5] G. Hughes, IEEE Transactions onInformation Theory 14 ,1968, 55.[6] M. Pal, P. Mather, International Journalof Remote Sensing 26 , 2005 ,1007.[7] K. Tan, P.J. Du, Spectroscopy andSpectral Analysis 28, 2008, 2009.[8] M. Chi, R. Feng, L. Bruzzone, Advancesin Space Research 41 (2008) 1793.[9] P.J. Du, X.M. Wang, K. Tan, G. Foody,8th International Symposium on SpatialAccuracy Assessment in NaturalResources and Environmental Sciences,25–27,June, 2008, Shanghai, WorldAcademic Union Ltd, Liverpool, 2008, p.138.[10] N. Ancona, R. Maglietta, E. Stella, PatternRecognition 39, 2006, 1588.[11] A. Plaza, J. Benediktsson, J. Boardman, J.Brazile, L. Bruzzone, G. Camps-Valls,J.Chanussot, M. Fauvel, P. Gamba, A.Gualtieri, Remote Sensing of Environment113, 2009, 110.[12] X. An, S.G. Su, T. Wang, S. Xu, W.J.Huang, L.D. Zhang, Spectroscopy andSpectral Analysis 27 ,2007, 1619.[13] C. Huang, J. Dun, Applied Soft ComputingJournal ,2007, 1381.[14] J. Inglada, ISPRS Journal ofPhotogrammetry and Remote Sensing 62,2007, 236.[15] L. Bruzzone, M. Chi, M. Marconcini,IEEE Transactions on Geoscience andRemote Sensing 44 , 2006, 3363.[16] G. Camps-Valls, L. Gomez-Chova, J.Calpe-Maravilla, J. Martin-Guerrero, E.Soria-Olivas, L. Alonso-Chorda, J.Moreno, IEEE Transactions onGeoscience and Remote Sensing 42 ,2004,1530[17] Helmi, Z., M., S. & , Affendi, S., &Shattri, M., The Performance ofMaximum Likelihood, Spectral AngleMapper, Neural Network and DecisionTree Classifiers in Hyperspectral ImageAnalysis,2007.[18] Richards, J., A., & Jia, X., Using SuitableNeighbors to Augment the Training Set inyperspectral Maximum LikelihoodClassification, 2008.[19] Hwang, J., T., Wavelet TextureExtraction and Image Classification ofHyperspectral Data Based on SupportVector Machine, 2009.[20] Mahesh Pal and Giles M. Foody, FeatureSelection for Classification ofHyperspectral Data by SVM, Member,IEEE TRANSACTIONS ONGEOSCIENCE AND REMOTE SENSING,VOL. 48, NO. 5, MAY 2010 2297[21] Anderson JR, Hardy EE, Roach JT andWitmer RE, A land use and land coverclassification system for use with remotesensor data, USGS Professional Paper964, Sioux Falls, South, 1976