文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 自然科学 > 天文/地理 > A Hybrid Object-Oriented Conditional Random Field
侵权投诉

A Hybrid Object-Oriented Conditional Random Field

遥感影像分类英文文献

AHybridObject-OrientedConditionalRandomField

Classi cationFrameworkforHighSpatialResolutionRemoteSensingImagery

YanfeiZhong,Member,IEEE,JiZhao,andLiangpeiZhang,SeniorMember,IEEE

Abstract—Highspatialresolution(HSR)remotesensingim-ageryprovidesabundantgeometricanddetailedinformation,whichisimportantforclassi cation.Inordertomakefulluseofthespatialcontextualinformation,object-orientedclassi cationandpairwiseconditionalrandom elds(CRFs)arewidelyused.However,thesegmentationscalechoiceisachallengingprobleminobject-orientedclassi cation,andtheclassi cationresultofpairwiseCRFalwayshasanoversmoothappearance.Inthispaper,ahybridobject-orientedCRFclassi cationframeworkforHSRimagery,namely,CRF+OO,isproposedtoaddresstheseproblemsbyintegratingobject-orientedclassi cationandCRFclassi cation.InCRF+OO,aprobabilisticpixelclassi cationis rstperformed,andthen,theclassi cationresultsoftwoCRFmodelswithdifferentpotentialfunctionsareusedtoobtainthesegmentationmapbyaconnected-componentlabelingalgorithm.Asaresult,anobject-levelclassi cationfusionschemecanbeused,whichintegratestheobject-orientedclassi cationsusingamajorityvotingstrategyattheobjectleveltoobtainthe nalclassi cationresult.Theexperimentalresultsusingtwomultispec-tralHSRimages(QuickBirdandIKONOS)andahyperspectralHSRimage(HYDICE)demonstratethattheproposedclassi -cationframeworkhasacompetitivequantitativeandqualitativeperformanceforHSRimageclassi cationwhencomparedwithotherstate-of-the-artclassi cationalgorithms.

IndexTerms—Classi cationfusion,conditionalrandom elds(CRFs),highspatialresolution(HSR),object-orientedclassi ca-tion,remotesensing.

I.INTRODUCTION

HEavailabilityofhighspatialresolution(HSR)remotesensingimageryobtainedfromsatellites(e.g.,IKONOS,QuickBird,andWorldView-2)increasesthepossibilityofaccu-rateEarthobservations[1].SuchHSRimageryprovidesvalu-ablegeometricanddetailedinformation,whichisimportantforapplicationssuchasdamageassessmentforenvironmentaldisasters,precisionagriculture,securityapplications,andurban

ManuscriptreceivedAugust21,2013;revisedNovember20,2013andJanuary7,2014;acceptedFebruary3,2014.DateofpublicationMarch11,2014;dateofcurrentversionMay30,2014.ThisworkwassupportedinpartbytheNationalBasicResearchProgramofChina(973Program)underGrant2011CB707105,bytheNationalNaturalScienceFoundationofChinaunderGrant41371344,andbytheFoundationfortheAuthorofNationalExcellentDoctoralDissertationofChinaunderGrant201052.

TheauthorsarewiththeStateKeyLaboratoryofInformationEngineer-inginSurveying,MappingandRemoteSensing,WuhanUniversity,Wuhan430079,China(e-mail:zhongyanfei@whu.edu.cn;zhaoji2015@gmail.com;zlp62@whu.edu.cn).

Colorversionsofoneormoreofthe guresinthispaperareavailableonlineatdoc.guandang.net.

DigitalObjectIdenti er10.1109/TGRS.2014.2306692

T

planning[2].Intheseapplications,mappingtheprede nedlandcovertypetothegroundareadenotedbyimagepixels(i.e.,theclassi cationtask)isparticularlyrelevant[3].

Thetraditionalpixel-basedclassi cationmethods,whicharealsocalledpixelwiseclassi cation,processeachpixelindependentlywithoutconsideringthecorrelationsbetweenneighboringpixels[4],[5].Theseclassi cationmethods,e.g.,supportvectormachine(SVM)[6],[7],neuralnetworks[8],andthemaximumlikelihoodclassi er(MLC)[9],havebeenwidelyusedforremotesensingimageryinmanyapplications.Oneparticulartypeofpixel-basedclassi cation,whichiscalledprobabilisticpixelwiseclassi cation(e.g.,MLC),canproducetheposteriorprobabilitiesofthesamplesbelongingtotheclasses,whichisparticularlyusefulinpracticalrecognitionsituations.Inaddition,certaindimensionreductionandman-ifoldlearning[10]–[14]methods,asapreprocessingstepofthepixel-basedclassi cation,canbe rstusedtoobtainthemorerepresentativefeatures,particularlyformediumorcoarsespatialresolutions.However,withthegradualincreaseinthespatialresolution,thespectralvariabilitywithineachlandcoverclassincreases,andtheseparabilitybetweenclassesdecreases.Asaresult,thesemethodsarelesseffectiveforHSRimagery,inwhichthesalt-and-pepperappearanceoftheclassi cationresultsinanincreaseinclassi cationerrors.

Inordertoovercomethisproblem,classi cationmethodsincorporatingspatialinformationintheclassi cationarenec-essarywhenconsideringthegeometricalinformationinHSRimagery[15].Therearetwomainapproachestoincorporatingspatialcontextualinformation:random eldsandobject-orientedclassi cationmethods.Theobject-orientedclassi ca-tionapproaches[16]processeachsegmentationregionasawholeratherthanindividualpixel,andthekeystepistoseg-menttheimageintorelativelyhomogeneousregionscalledseg-mentationregionsorobjects[17].Manydifferentapproaches,suchasthefractalnetevolutionapproach(FNEA)[18],themean-shiftsegmentation(MSS)approach[19],andwatershedsegmentation[20],havebeenusedtodealwiththisproblem.Amajorityvotingstrategyafterthepixel-basedclassi cationorthedirectclassi cationusingregionfeaturesisthenappliedtotransformthesegmentationmapintotheclassi cationresult[21]–[23].Usingthesegmentationregionsasthebasicanalysisunitshelpstoovercomethesalt-and-pepperappearanceoftheclassi cation.Thereasonforthisisthatincorporatingthespatialcontextualinformation,suchassize,shape,texture,andgeometricalstructure,helpstoalleviatethewithin-class

0196-2892©2014IEEE.Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.

Seedoc.guandang.net.

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email