资源描述:
《遥感图像最大似然分类方法的EM 改进算法》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、第31卷 第3期测 绘 学 报Vol.31,No.32002年8月ACTAGEODAETICAetCARTOGRAPHICASINICAAug.,2002 文章编号:100121595(2002)0320234206中图分类号:P237文献标识码:A遥感图像最大似然分类方法的EM改进算法12314骆剑承,王钦敏,马江洪,周成虎,梁 怡(1.中国科学院地理科学与资源研究所,北京100101;2.福州大学,福建福州350002;3.西安交通大学,陕西西安710049;4.香港中文大学地理系,香港)TheEM2basedMaximumLikelihoodClassifierforRemotely
2、SensedData12314LUOJian2cheng,WANGQin2min,MAJiang2hong,ZHOUCheng2hu,LEUNGYee(1.InstituteofGeographicalSciencesandNaturalResourcesResearch,CAS,Beijing100101,China;2.FuzhouUniversity,Fuzhou350002,China;3.Xi’anJiaotongUniversity,Xi’an710049,China;4.DepartmentofGeography,theChineseUniversityofHongKong,H
3、ongKong,China)Abstract:Basedonparametricdensitydistributionmodel,themaximumlikelihoodclassification(MLC)mightbeoneofthemostpopularmethodsforremotesensingimageclassification.Bycomparisonwithnon2parametricapproaches,MLChasseveraldistinctadvantages,suchasitsclearparametricinterpretability,feasibleinte
4、grationwithpriorknowledgebasedonBayesiantheory,andrelativesimplerealization,etc.However,remotesensinginformationhassomedegreeofdefinitesta2tisticalcharacteristic,butaswellasholdsthehighrandomnessandcomplexity,whichgenerallybehavesasmixturedensitydistributioninfeaturespace.Ifthedistributionsofcertai
5、ncategoriesinfeaturespacearesodiscretethattheymightnotobeytothesingleassumeddistribution,orthetrainingsamplesarenotsufficientenoughthattheycannotrepresenttheoveralldistributions,itoftenbringsongreatbiasbetweenobtainedresultsandpracticalsituations.Inthisarticle,wefirstlyintro2duceintotheexpectationm
6、aximization(EM)algorithminordertoextendtheconventionalMLCapproachtomixturedensitymodel.EMassumesthattheoveralldistributioncouldbedecomposedintoinfiniteparametricdistributions.Themodelshouldbefirstlyassumedthatwholedistributioncouldbeseparatedintoinfiniteparametricdensitydistributions,thenbyEMiterat
7、ivecomputationthemaximumlikelihoodparametersofeachproportionaldistributioncanbeestimated.BetterparameterestimatescanbeobtainedbyexploitingalargenumberofunlabeledsamplesinadditiontotrainingsamplesusingtheEMa