Research Article A Method of Spatial Mapping and Reclassification for High-Spatial-Resolution Remote Sensing Image Classification

Transcription

1 Hindwi Pulishing Corportion The Scientific World Journl Volume 2013, Article ID , 7 pges Reserch Article A Method of Sptil Mpping nd Reclssifiction for High-Sptil-Resolution Remote Sensing Imge Clssifiction Guizhou Wng, 1,2 Jino Liu, 1 nd Guojin He 1 1 Institute of Remote Sensing nd Digitl Erth, Chinese Acdemy of Sciences, Beijing , Chin 2 University of Chinese Acdemy of Sciences, Beijing , Chin Correspondence should e ddressed to Guojin He; gjhe@ceode.c.cn Received 24 Septemer 2013; Accepted 19 Novemer 2013 Acdemic Editors: Z. Hou nd R. D. J. Romero-Troncoso Copyright 2013 Guizhou Wng et l. This is n open ccess rticle distriuted under the Cretive Commons Attriution License, which permits unrestricted use, distriution, nd reproduction in ny medium, provided the originl work is properly cited. This pper presents new clssifiction method for high-sptil-resolution remote sensing imges sed on strtegic mechnism of sptil mpping nd reclssifiction. The proposed method includes four steps. First, the multispectrl imge is clssified y trditionl pixel-sed clssifiction method (support vector mchine). Second, the pnchromtic imge is sudivided y wtershed segmenttion. Third, the pixel-sed multispectrl imge clssifiction result is mpped to the pnchromtic segmenttion result sed on sptil mpping mechnism nd the re dominnt principle. During the mpping process, n re proportion threshold is set, nd the regionl property is defined s unclssified if the mximum re proportion does not surpss the threshold. Finlly, unclssified regions re reclssified sed on spectrl informtion using the minimum distnce to men lgorithm. Experimentl results show tht the clssifiction method for high-sptil-resolution remote sensing imges sed on the sptil mpping mechnism nd reclssifiction strtegy cn mke use of oth pnchromtic nd multispectrl informtion, integrte the pixelnd oject-sed clssifiction methods, nd improve clssifiction ccurcy. 1. Introduction With the development of improved sensors nd powerful computtion technology, high-sptil-resolution remote sensingdthveeenmoreesilycquiredndwidely pplied [1]. High-sptil-resolution remote sensing imges contin more informtion nd hve incresed the detil t which erth oservtions cn e mde. The undnce of informtion, on one hnd, hs promoted the ppliction of remote sensing methods ut, on the other, rings new technologicl chllenges to the dt nlysis. One chllenge is tht trditionl imge clssifiction technology cn no longer stisfy the needs of high-sptil-resolution remote sensing imge clssifiction. High-sptil-resolution remote sensing imgery, such s SPOT-5,IKONOS,ndQuickBird,hseenusedinmny fields in recent yers [2]. They hve een pplied for urn plnning, urn chnge detection, tree cnopy mpping, ecologicl environment monitor, precision griculture, nd so forth [3]. The min difference etween high-sptilresolution remote sensing imge nd low- or mediumresolution remote sensing imge is tht the high-sptilresolution imge provides more useful informtion, such s shpe nd texture. Therefore, the extrction of geogrphicl informtion from high-sptil-resolution stellite imge is topicl [2]. The trditionl method of clssifiction for high-sptilresolution imges hs een proven to hve severl drwcks, such s low clssifiction ccurcy, the derivtion of very limited sptil informtion, nd slt nd pepper effects [2]. Therefore, novel nd efficient nlysis techniques re needed for processing nd nlyzing of high-sptilresolution remote sensing imges. Mny studies hve een done on segmenttion nd clssifiction of high-sptilresolution remote sensing imges [2 8]. Trlk et l. [4] presentednewspectrl-sptilclssifictionscheme

2 2 The Scientific World Journl for hyperspectrl imges, comining the pixel-sed support vector mchine clssifiction results nd the wtershed segmenttion regions together. Bruzzone nd Crlin [6] proposed novel pixel-sed system for the clssifiction of high-sptil-resolution imges. The sptil context informtion of ech pixel extrcted from multilevel segmenttion imges ws used to otin more ccurte nd relile clssifiction result. Chen et l. [2] introduced modified oject-oriented clssifiction lgorithm integrting multichrcteristics of high-sptil-resolution remote sensing imge. Ünsln nd Boyer [7] put forwrd new clssifiction method of lnd development in high resolution pnchromtic stellite imges using stright line sttistics. Slehi et l. [8] developed hierrchicl rule-sed ojectsed clssifiction frmework coupled with height points for complex urn environment clssifiction. The rule set ws extrcted from trining set of QuickBird imge coupled with lyer of height points. In ddition, the morphologicl sed segmenttion nd clssifiction ws investigted in the work of Dll Mur et l. [5] ndpesresindbenediktsson [9]. Although some improved lgorithms cn increse the clssifiction ccurcy y mking use of spectrl nd texturl informtion, the clssifiction results still cnnot stisfy ll ctul needs. Most of these methods were developed on fusion imges or multispectrl imges. The sptil reltionship of high-sptil-resolution remote sensing imge etween pnchromtic nd multispectrl nds hs not een fully considered. This pper proposes new high-sptil-resolution remote sensing imge clssifiction method sed on mechnism of sptil mpping nd strtegy of reclssifiction. The nlysis mkes use of oth spectrl nd sptil informtion from high-sptil-resolution remote sensing dt. The clssifiction frmework uses sptil mpping mechnism to fit the specil dt formt nd the content of high-sptilresolution remote sensing imges. This lgorithm is not sed on fusion imge, ut rther on rw high-sptil-resolution remote sensing dt, which mkes full use of the sptil resolution reltionship of pnchromtic nd multispectrl imges. QuickBird nd SPOT-5 stellite dt were employed in series of experiments nd comprtive nlyses. Trditionl pixel-sed SVM, oject-oriented SVM, nd the method proposed in [4] (SVM + Mjority Voting, SVMMV) were used for comprison to nlyze the dvntge of the proposed method. Experimentl results show tht clssifiction method sed on sptil mpping mechnism nd reclssifiction strtegy for high-sptil-resolution remote sensing dt cn mke full use of the informtion in oth pnchromtic nd multispectrl nds, integrte the pixelndoject-sedclssifictionmethods,ndimprovethe clssifiction ccurcy. 2. Methodology This section detils the support vector mchine clssifiction method, nd wtershed imge segmenttion, nd outlines the sptil mpping mechnism nd reclssifiction strtegy. A Multispectrl imge Pixel-sed clssifiction Mpping clssifiction result from multispectrl to pnchromtic Reclssify unclssified regions Finl pnchromtic clssifiction result Pnchromtic imge Imge segmenttion Figure 1: The flow chrt of the proposed clssifiction method. flow chrt of the proposed clssifiction method is shown in Figure Support Vector Mchine Clssifiction. The first step in the proposed method is the pixel-sed clssifiction of multispectrl imge. There re mny possile imge clssifiction lgorithms for remote sensing imges. Ech lgorithm hs its unique dvntges nd weknesses. In this pper, we focus on the ppliction of the sptil mpping mechnism nd reclssifiction strtegy for high-sptilresolution remote sensing imges. In order to improve the overll efficiency of pproch nd to simplify it, support vector mchine clssifier is chosen. Theoreticlly, ny pixelsed clssifiction method cn e pplied in the proposed method. Support vector mchine (SVM) is supervised nonprmetric sttisticl lerning technique nd widely used in clssifiction of remote sensing imges [10]. SVM is well t solving nonliner, high dimensionl, nd limited trining smples [8]. In this pper, LIBSVM lirry is used to implement the SVM clssifiction lgorithm [11] Wtershed Imge Segmenttion. Imge segmenttion is n importnt prt of imge interprettion, especilly for high-sptil-resolution remote sensing imges [12, 13]. High-sptil-resolution remote sensing imges contin more informtion of ground ojects nd show gret diversity of them. The purpose of segmenttion is to divide n imge into homogeneous regions. Wtershed trnsformtion is powerful mthemticl morphology technique for imge segmenttion [4, 14]. A wtershed lgorithm is good choice for high-sptil-resolution remote sensing imges ecuse of its fst segmenttion speed [15]. In this pper, leled wtershed segmenttion lgorithm wsusedtosegmentthepnchromticimge.morphologicl opertors were used to chrcterize the grdient of the imge, which enled the wtershed segmenttion lgorithm to lel res within the imge.

3 The Scientific World Journl Sptil Mpping Mechnism. High-sptil-resolution remote sensing dt often contin two types of imges, those with single pnchromtic nd nd those with four multispectrl nds [16]. For exmple, QuickBird imges hve pnchromtic nd with resolution of 0.6 m nd four multispectrl nds with resolution of 2.4 m. A pnchromtic imge of high-sptil-resolution remote sensing dt contins most of the sptil informtion, wheres multispectrl imge hs most of the spectrl informtion. In order to mke full use of pnchromtic nd multispectrl imge informtion, imge fusion is widely used to integrte the two. The fusion lgorithm merges the highresolution pnchromtic nd low-resolution multispectrl imgery to crete n enhnced high-resolution multispectrlimge.then,thefusedimgeisusedinsusequent pplictions. Note tht the fused imge is n estimtion which my cuse spectrl distortion nd ffect the ccurcy of clssifiction results [16]. The effect of fusion directly determines the susequent ppliction ccurcy. The proposed clssifiction frmework uses sptil mpping mechnism to mke full use of sptil nd spectrl informtion in high-sptil-resolution remote sensing dt. In the presented method, the rw high-sptil-resolution dt, insted of the fusion imge, ws directly clssified sed on sptil mpping mechnism nd reclssifiction strtegy. For exmple, Figure 2 shows the sptil mpping reltionship etween pnchromtic nd multispectrl imges, tking resolution rtio of 1 : 4. One pixel in the multispectrl imge corresponds to sixteen pixels in the pnchromtic imge. For exmple, if pixel hs position (x i,y j ) in the pnchromtic imge, then i=1,2,...,m; j=1,2,...,n;thecorresponding position of the pixel in the multispectrl imge is (x i,y j ), where x i = Int(x i / ) nd y j = Int(y j/ ). The Int()function rounds numer to the nerest integer. The pixel-sed multispectrl clssifiction result ws mpped to the pnchromtic segmenttion result sed onthesptilmppingmechnismnd redominnt principle. When using the re dominnt principle s mpping ruler, the rel proportion of ech clss in ech region is computed nd clss lel corresponding to the mximum re proportion is ssigned to tht region. Figure 3 shows the sptil mpping mechnism sed on the re dominnt principle. On the left is the pnchromtic imge with the originl clss lels, wheres sptil mpping is sedonthemultispectrlimgepixel-sedclssifiction result. On the right re the regions of the pnchromtic imge segmenttion. After sptil mpping sed on the re dominnt principle, region one is leled nd region two is leled. The re proportion of ech clss represents the region s memership to every clss. Regions composed of only one clss of pixels hve higher re proportion close to one for this clss nd zero for other clsses. However, regions composed of pixels elonging to severl different clsses hve lower re proportion for every clss. The mximum re proportion of ech region reflects the miguousness mpping from the pixel-sed clssifiction. During the mpping process, n re proportion threshold (0 <T 1) Multispectrl imge Pnchromtic imge Figure 2: The sptil mpping reltionship etween multispectrl nd pnchromtic imges, tking resolution rte of 1 : 4 for exmple. is set. The region is leled s unclssified if the mximum re proportion does not surpss the threshold. The greter the threshold, the greter the clssified regions reliility. In this pper, the threshold ws set to 0.6. The regionl property is considered very relile if the mximum re proportion is greter thn 0.6. Unclssified regions re reclssified in the next step Reclssifiction Strtegy. Unclssified regions in the imges will e reclssified sed on spectrl informtion usingtheminimumdistncetomen(mdtm)lgorithm [17]. The minimum distnce clssifiction uses men vector for ech clss nd clcultes the Eucliden distnce from ech unclssified region to the men vector for ech clss. All unclssified regions were clssified to the closest clss. In this pper, the clssified regions through sptil mpping mechnism sed on the re dominnt principle were used s trining smples. The regionl spectrl vector is clculted y the men spectrl vector of pixels contined in ech region. The spectrl vector of ech pixel in pnchromtic imge is otined from multispectrl imge y sptil mpping mechnism. A MDTM clssifier computes the Euclidin distnce in spectrl spce etween the men of every clss in the trining set nd the region to e clssified. The Euclidin distnce etween the men of clss nd n unclssified region in the n dimensionl spectrl feture spce is given s [18] D ED =( n i=1 (x i C i ) 2 ) 1/2, (1) where n is the dimensionlity of dt, x i is the men spectrl vlue of the ith nd of the unclssified region, nd C i is the men spectrl vlue of the ith nd of one clss. The unclssified region is then ssigned to the clss where D ED is miniml.

4 4 The Scientific World Journl Originl clss lel Segmenttion regions () () Figure 3: Mpping mechnism y the re dominnt principle. () () (c) Rod Wter Artificil grss Building Bre lnd Plstic trck Vegettion Shdow Unclssified (d) (e) (f) Figure 4: QuickBird imges nd clssifiction results. () is the fused pseudocolor synthetic imge, () is the primitive mpped clssifiction imge, (c) shows the finl clssifiction result otined y the proposed method, nd (d) is result of the pixel-sed SVM. The clssifiction result of the oject-oriented SVM is shown in (e), nd the SVMMV is shown in (f). 3. Experimentl Results nd Anlysis To evlute the performnce of the proposed clssifiction pproch, two susets of high-sptil-resolution remote sensing imges, QuickBird nd SPOT-5 stellite imges, were used in series of experiments nd comprtive nlyses. To nlyze the dvntges of the proposed method for high-sptil-resolution remote sensing imges, trditionl pixel-sed SVM, oject-oriented SVM, nd the method proposed in [4] (SVM + Mjority Voting, SVMMV) on fusion imges were used for comprison. The pnchromtic nd multispectrl imges were fused y the PANSHARP method in PCI softwre. Leled wtershed trnsformtion ws pplied to the morphology grdient of pnchromtic imge to otin segmenttion regions. The oject-oriented SVM ws pplied on the pnchromtic segmenttion regions nd the region feture ws computed from the fusion imges. The SVMMV method presented spectrl-sptil clssifiction

5 The Scientific World Journl 5 () () (c) Forest lnd Wter Residentil lnd Bre lnd Industril lnd Agriculturl lnd Rod Unclssified (d) (e) (f) Figure 5: SPOT-5 imges nd clssifiction results. () is the fused pseudocolor synthetic imge, () is the primitive mpped clssifiction imge, (c) shows the finl clssifiction result otined y the proposed method, nd (d) is result of the pixel-sed SVM. The clssifiction result of the oject-oriented SVM is shown in (e), nd the SVMMV is shown in (f). Tle 1: Optiml prmeters of ll SVM clssifiction experiments. Optiml prmeters QuickBird dtset SPOT-5 dtset C γ C γ Pixel-sed SVM Oject-oriented SVM SVMMV The proposed method scheme, comining the pixel-sed support vector mchine clssifiction results nd the wtershed segmenttion regions through mjority voting. The SVM clssifier with Gussin rdil sis function (RBF) kernel ws pplied in ll experiments. The optiml prmeters C (prmeter tht controls the mount of penlty during the SVM optimiztion) nd γ (prmeter tht descries the spred of the RBF kernel) were chosen y fivefold cross vlidtion [11, 19]. Tle 1 reports the optiml prmeters of ll SVM clssifiction experiments. To ssess clssifiction ccurcy, confusion mtrix is used. Confusion mtrices re otined y selecting points with strtified rndom smpling nd ssessing the clss of ech point s clculted y ech of the four methods. The reference clssifiction imges were generted through precise mnul interprettion on fusion imges. The producer, user, nd overll clssifiction ccurcies re clculted from the confusion mtrix. The first clssifiction experiment ws performed with the QuickBird imge, including the pnchromtic nd with resolution of 0.6 m nd four multispectrl nds with resolution of 2.4 m. The size of multispectrl imge is , wheres tht of the pnchromtic imge is The optiml prmeters of SVM clssifiers on QuickBird dtset re shown in Tle 1. The clssifiction results re shown in Figure 4. Tle 2 shows the producer, user, nd overll clssifiction ccurcies for the QuickBird Imge clssifiction. A second clssifiction experiment ws performed with the SPOT-5 stellite imge, consisting of pnchromtic nd with resolution of 2.5 m nd four multispectrl nds with resolution of 10.0 m. The size of the multispectrl imge is , wheresitis for the pnchromtic imge. The optiml prmeters of SVM clssifiers on SPOT- 5 dtset re shown in Tle 1. The results of clssifiction re shown in Figure 5. Tle 3 shows the producer, user, nd overll clssifiction ccurcies for the SPOT-5 imge clssifiction. By compring the ccurcies of the vrious clssifictions, the proposed method sed on sptil mpping mechnism nd strtegy of reclssifiction cn e seen to otin etter clssifiction results thn pixel-sed SVM, oject-oriented SVM, nd SVMMV. Tles 2 nd 3 show tht the overll

6 6 The Scientific World Journl Tle 2: Producer nd user clssifiction ccurcies for QuickBird imge clssifiction. Proposed method Pixel-sed SVM Oject-oriented SVM SVMMV Clss ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % Rod Building Vegettion Wter Bre lnd Shdow Artificil grss Plstic trck Averge ccurcy Overll ccurcy Tle 3: Producer nd user clssifiction ccurcies for SPOT-5 imge clssifiction. Proposed method Pixel-sed SVM Oject-oriented SVM SVMMV Clss ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % ccurcy % Forest lnd Residentil lnd Bre lnd Industril lnd Wter Agriculturl lnd Rod Averge ccurcy Overll ccurcy ccurcies of the proposed method re higher thn those of the pixel-sed SVM, oject-oriented SVM, nd SVMMV methods. The ccurcies of the pixel-sed SVM results were lower thn other three methods, ecuse the technique suffers from slt-nd-pepper effects. Some of the noise in the pixel-sed clssifiction result cn e reduced through postprocessing (e.g., mjority filtering), ut postprocessing cn result in the disloction of clss oundries nd influence the outcome of susequent pplictions. The SVMMV method, which comined the pixel-sed SVM clssifiction results nd the wtershed segmenttion regions through mjority voting, got higher clssifiction ccurcies thn pixel-sed SVM clssifiction. The slt-nd-pepper effects hd een reduced nd more homogeneous regions were otined in the SVMMV clssifiction mps. The proposed method otined higher ccurcies thn the SVMMV method, ecuse it mde full use of the sptil reltionship etween pnchromtic nd multispectrl imges nd strtegy of reclssifiction. Although the oject-oriented SVM, pixel-sed SVM, nd SVMMV clssifiction results were otined from veryhigh-sptil-resolution fusion imge, the imge ws nevertheless n estimtion tht could introduce spectrl distortion nd confusion. In order to mke further comprtive nlysis of the clssifiction ccurcy, the re percents of ech clss for the four methods on the QuickBird dtset were computed (Figure 6). Note tht the clss re percents of the proposed method re the closest to the rel re percents. According to the experimentl results, the proposed clssifiction method sed on sptil mpping mechnism nd reclssifiction strtegy cn otin higher ccurcy thn the pixel-sed SVM, oject-oriented SVM, nd SVMMV clssifictionmethods.theproposedmethodcnmkefull use of the informtion oth in pnchromtic nd multispectrl nds nd integrte the pixel-sed nd oject-sed clssifiction methods. In this pper, only spectrl fetures in the imges were pplied to the clssifiction process. Further work is required to integrte texturl fetures. 4. Conclusions A new high-sptil-resolution remote sensing imge clssifiction method sed on sptil mpping mechnism nd reclssifiction strtegy hs een presented in this pper,

7 Rod Wter Shdow Artificil grss The Scientific World Journl 7 Are (%) Building Vegettion Rel re (%) Proposed method (%) Pixel-sed SVM (%) Bre lnd Plstic trck Oject-oriented SVM (%) SVMMV (%) Figure 6: Are sttistics of QuickBird dt clssifiction results. in which pixel-sed clssifiction method nd n ojectsed segmenttion nd clssifiction method were integrted y sptil mpping mechnism nd reclssifiction strtegy. Furthermore, the proposed method ws pplied on rw high-sptil-resolution remote sensing dt insted of fusion imges. Experimentl results hve demonstrted tht theproposedmethodcnmkefulluseoftheinformtion in oth pnchromtic nd multispectrl nds, integrte the pixel-sed nd oject-sed segmenttion nd clssifiction methods, nd otin higher finl clssifiction ccurcy. Acknowledgments This work ws sponsored y the Ntionl Nturl Science Foundtion of Chin under the Grnt no , the Ntionl Key Technology Support Progrm of Chin (2012BAH27B05), nd the Ntionl Ecologicl Environment Specil Project (STSN-10-03). The uthors wish to thnk the nonymous reviewers who provided constructive comments tht improved the qulity nd clrity of the pper. References [1] J. Yun nd G. He, A new clssifiction lgorithm for high sptil resolution remote sensing dt, in Proceedings of the Interntionl Conference on Erth Oservtion Dt Processing nd Anlysis (ICEODPA 08), vol. 7285, Wuhn, Chin, Decemer [2] Z. Chen, G. Wng, nd J. Liu, A modified oject-oriented clssifiction lgorithm nd its ppliction in high-resolution remote-sensing imgery, Interntionl Journl of Remote Sensing,vol.33,no.10,pp ,2012. [3] T. Novck, T. Esch, H. Kux, nd U. Still, Mchine lerning comprison etween WorldView-2 nd QuickBird-2-simulted imgery regrding oject-sed urn lnd cover clssifiction, Remote Sensing,vol.3,no.10,pp ,2011. [4]Y.Trlk,J.Chnussot,J.A.Benediktsson,J.Angulo,nd M. Fuvel, Segmenttion nd clssifiction of hyperspectrl dt using wtershed, in Proceedings of the IEEE Interntionl Geoscience nd Remote Sensing Symposium, pp. III652 III655, Boston, Mss, USA, July [5]M.DllMur,J.A.Benediktsson,B.Wske,ndL.Bruzzone, Morphologicl ttriute profiles for the nlysis of very high resolution imges, IEEE Trnsctions on Geoscience nd Remote Sensing,vol.48,no.10,pp ,2010. [6] L. Bruzzone nd L. Crlin, A multilevel context-sed system for clssifiction of very high sptil resolution imges, IEEE Trnsctions on Geoscience nd Remote Sensing, vol.44,no.9, pp , [7] C. Ünsln nd K. L. Boyer, Clssifying lnd development in high-resolution pnchromtic stellite imges using strightline sttistics, IEEE Trnsctions on Geoscience nd Remote Sensing,vol.42,no.4,pp ,2004. [8] B. Slehi, Y. Zhng, M. Zhong, nd V. Dey, Oject-sed clssifiction of urn res using VHR imgery nd height points ncillry dt, Remote Sensing, vol. 4, no. 8, pp , [9] M. Pesresi nd J. A. Benediktsson, A new pproch for the morphologicl segmenttion of high-resolution stellite imgery, IEEE Trnsctions on Geoscience nd Remote Sensing, vol.39,no.2,pp ,2001. [10] G. Mountrkis, J. Im, nd C. Ogole, Support vector mchines in remote sensing: review, ISPRS Journl of Photogrmmetry nd Remote Sensing,vol.66,no.3,pp ,2011. [11] C.-C. Chng nd C.-J. Lin, LIBSVM: lirry for support vector mchines, ACM Trnsctions on Intelligent Systems nd Technology,vol.2,no.3,rticle27,2011. [12] G. Fu, H. Zho, C. Li, nd L. Shi, Segmenttion for highresolution opticl remote sensing imgery using improved qudtree nd region djcency grph technique, Remote Sensing,vol.5,no.7,pp ,2013. [13] C.Zheng,L.Wng,R.Chen,ndX.Chen, Imgesegmenttion using multiregion-resolution MRF model, IEEE Geoscience nd Remote Sensing Letters, vol. 10, no. 4, pp , [14] Y. Trlk, J. Chnussot, nd J. A. Benediktsson, Segmenttion nd clssifiction of hyperspectrl imges using wtershed trnsformtion, Pttern Recognition, vol.43,pp , [15]G.WngndG.He, Humnvisulsystemsedprocessing for high resolution remote sensing imge segmenttion, in Proceedings of the 2nd Interntionl Conference on Signl Processing Systems (ICSPS 10), pp. V1474 V1478, Dlin, Chin, July [16] G. Wng, G. He, nd J. Liu, A new clssifiction method for high sptil resolution remote sensing imge sed on mpping mechnism, in Proceedings of the Interntionl Conference on Geogrphic Oject-Bsed Imge Anlysis (GEOBIA 12),pp , Rio de Jneiro, Brzil, My [17] A. G. Wcker nd D. A. Lndgree, Minimum distnce clssifiction in remote sensing, Tech. Rep. 25, LARS, [18] A. K. Shckelford nd C. H. Dvis, A hierrchicl fuzzy clssifiction pproch for high-resolution multispectrl dt over urn res, IEEE Trnsctions on Geoscience nd Remote Sensing,vol.41,no.9,pp ,2003. [19] Y. Trlk, J. Chnussot, nd J. A. Benediktsson, Segmenttion nd clssifiction of hyperspectrl imges using minimum spnning forest grown from utomticlly selected mrkers, IEEE Trnsctions on Systems, Mn, nd Cyernetics B,vol.40, no.5,pp ,2010.

8 Interntionl Journl of Rotting Mchinery Engineering Journl of Hindwi Pulishing Corportion The Scientific World Journl Hindwi Pulishing Corportion Interntionl Journl of Distriuted Sensor Networks Journl of Sensors Hindwi Pulishing Corportion Hindwi Pulishing Corportion Hindwi Pulishing Corportion Journl of Control Science nd Engineering Advnces in Civil Engineering Hindwi Pulishing Corportion Hindwi Pulishing Corportion Sumit your mnuscripts t Journl of Journl of Electricl nd Computer Engineering Rootics Hindwi Pulishing Corportion Hindwi Pulishing Corportion VLSI Design Advnces in OptoElectronics Interntionl Journl of Nvigtion nd Oservtion Hindwi Pulishing Corportion Hindwi Pulishing Corportion Hindwi Pulishing Corportion Chemicl Engineering Hindwi Pulishing Corportion Active nd Pssive Electronic Components Antenns nd Propgtion Hindwi Pulishing Corportion Aerospce Engineering Hindwi Pulishing Corportion Hindwi Pulishing Corportion Interntionl Journl of Interntionl Journl of Interntionl Journl of Modelling & Simultion in Engineering Hindwi Pulishing Corportion Shock nd Virtion Hindwi Pulishing Corportion Advnces in Acoustics nd Virtion Hindwi Pulishing Corportion