基于半監(jiān)督集成的遙感圖像的分割和分類.pdf

1、西安電子科技大學(xué)碩士學(xué)位論文基于半監(jiān)督集成的遙感圖像的分割和分類姓名：王婷申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：模式識(shí)別與智能系統(tǒng)指導(dǎo)教師：焦李成20091201ABSTRACTIIIABSTRACTTheearthobservationtechnologyofremotesensinghasbeenwidelyusedinbattlefieldintelligenceandreconnaissance，targetrecognition，clima

2、techangemonitoring，vegetationmonitoring，disastermonitoring，digitalearthetcSegmentation，classificationandrecognitionarethekeytechnologiesinremotesensingimageprocessingandinterpretationInrecentyears，semi—supervisedlearning

3、hasattractedincreasingattention，SOthispaperhavearesearchonremotesensingimageprocessingusingthesemi—supervisedlearningtheoryandmethodsThemaincontributionsofthispapercallbesummarizedasfollows：1Asemi—supervisedspectralclust

4、eringalgorithmforlargescaledataclusteringisproposedTheuniformsamplingtechniquecombinedwithnearestneighborruleisusedforthelargescaledataclusteringPairwiseconstraintsareincorporatedintospectralclustertoconstructsemi—superv

5、isedspectralclusteringinwhichtheselftuningparameterisappliedtoavoidtheselectionofthescalingparameterTheexperimentalresultsonthesegmentationofmultispectralremotesensingimageshowtheeffectivenessandpotentialoftheproposedmet

6、hod2BasedontheSemi—BoostensemblemethodwithsemisupervisedpathbasedsimilarityatargetrecognitionsystemfortheSARimagesisputforwardAddthepairwiseconstraintsinformationgeneratedbythelabeledsamplesintothesimilaritymatrixAndguid

7、edbythesimilaritymatrixtheunlabeledsamples、析tllllighconfidenceareaddedintothetrainingsamplesandfinallytheperformanceofthealgorithmisimprovedbyensemblingmulti—SVMswithweightedcomponentTheexperimentalresultsshowthatthetarg

8、etrecognitionaccuracyoftheproposedalgorithmishigherthanthatofothertraditionalmethods3AnspectralclusteringensemblemethodbasedonboostingisproposedandappliedtothesegmentationofSARimagesTheindividualspectralclusteringalgorit

9、hmisconstructedbyadaptiveresamplingtechniquebasedonboostingandrandomscalingparametersThefinalclusteringisobtainedbyMCLAforthecombinationofcomponentclusteringTheexperimentalresultsonUCIdatasets，thesegmentationofsynthetict