Large-scaleSatelliteImageBrowsingusingAutomaticSemanticCategorization
andContent-basedRetrieval
AshishParulekar RitendraDatta JiaLi§JamesZ.Wang¶
ThePennsylvaniaStateUniversity,UniversityPark,PA,USA
{ashishp,datta,jiali,jwang}@psu.edu
Abstract
Weapproachtheproblemoflarge-scalesatelliteimagebrowsingfromacontent-basedretrievalandsemanticcategorizationperspective.Atwo-stagemethodforquerybasedautomaticretrievalofsatelliteimagepatchesisproposed.Thesemanticcategoryofquerypatchesaredeterminedandpatchesfromthatcategoryarerankedbasedonanimagesimilaritymeasure.Semanticcategorizationisdonebyalearningapproachinvolvingthetwo-dimensionalmulti-resolutionhiddenMarkovmodel(2-DMHMM).Patchesthatdonotbelongtoanytrainedcategoryarehandledusingasupportvectormachine(SVM)basedclassi er.Experimentsyieldpromisingresultsinmodelingsemanticcategorieswithinsatelliteimagesusing2-DMHMM,producingaccurateandconvenientbrowsing.Wealsoshowthatpriorsemanticcategorizationimprovesretrievalperformance.
Sciences,EnvironmentalSciencesandinservingothercivilianandmilitarypurposes.
Overthelastthreedecades,therehasbeenanoverwhelmingamountofworkdoneinthe eldofautomatedclassi cationofsatelliteimages,oneoftheearliestbeing[7]byHaralicketal.Inthemorerecentyears,manydifferenttechniqueshavebeenproposed,someofwhichincludetextureanalysis[10,8],Markovrandom elds[13],geneticalgorithms[14],fuzzyapproaches[6,18],bayesianclassi ers[11,5]andneuralnetworks[2,1].Someothershaveapproachedsatelliteimageclassi cationasanapplicationofcontent-basedimageretrieval[10,12].Inarecentsurvey[17]onsatelliteimageclassi cationresultspublishedinthelast fteenyears,ithasbeenreportedthattheclassi cationaccuracyhasnotincreasedsigni cantlyovertheyears.Thepaperreportsameanoverallpixel-wiseclassi cationaccuracyacrossalltheseexperimentalresultsat76.19%withastandarddeviationof15.59%.However,assuggestedbytheauthor,theseresultsneedtobetakenwithcaution,sinceanumberofexperimentalparametershavevariedovertheyearsandacrossdifferentapproachestowardsclassi cation.Nonetheless,webelievethatthisiscertainlyindicativeoftheexistenceofanupperboundontheclassi cationaccuracyinthistypeofimagery,giventhecurrenttechnology.Asaresult,thispaperismoreaboutapproachestowardsafastandscalableworkingsystemwhilemaintainingrespectableclassi cationaccuracy.Ourworkinvolveslearningthesemanticsofsatelliteimagepatchesforquerying,retrievalandbrowsing,usingacontent-basedimageretrievalapproach.Thepurposeisnotsomuchtoimproveupontheaccuracyratesbutrathertoshowthat2-DMHMMisaneffectiveandef cientwaytojointlymodelthetextureandspatialstructureofsemanticcategorieswithinsatelliteimagery.Moreover,weshowthroughexperimentsthatperformingcategorizationpriortoapplyingimageretrievaltechniquesincreasesspeedandaccuracyinbrowsinglargedatabasesofsatelliteimagery.Ratherthangeneratinganautomaticsegmentationoftheterrainatpixellevel,weintendtoprovideabrowsingplatformthroughpatch-levelclassi cationfollowedbya1
1.Introduction
Today,theneedforreliable,automated,satelliteimageclassi cationandbrowsingsystemsismorethaneverbefore.Everydaythereisamassiveamountofremotely-senseddatabeingcollectedandsentbyterrestrialsatellitesforanalysis.Theuseofautomatedtoolsforthisanalysishasbecomeimperativeduetothelargevolumesrequiredtobeprocessedonadailybasis.Applicationsofthisclassi cationlieindiverse eldssuchasGeography,Geology,Archaeology,Atmospheric
materialisbaseduponworksupportedbytheUSNational
ScienceFoundationunderGrantNos.IIS-0219272,EIA-0202007,andIIS-0347148,thePNCFoundation,andSUNMicrosystems.DiscussionswithsatelliteimageanalystBradleyFarsterwerehelpfulinselectingthetrainingsamplesandacquiringthedataset.Riemann.ist.psu.eduprovidesmoreinformation.
A.Parulekarisaf liatedwiththeDepartmentofElectricalEngineering.
R.Dattaisaf liatedwiththeDepartmentofComputerScienceandEngineering.
§J.Liisaf liatedwiththeDepartmentofStatisticsandtheDepartmentofComputerScienceandEngineering.
¶J.Z.Wangisaf liatedwiththeSchoolofInformationSciencesandTechnologyandtheDepartmentofComputerScienceandEngineering.
The
SemreX: Towards Large-Scale Literature Information Retrieval and Browsing with Semantic Association* Xiaomin Ning, Hai Jin, Hao Wu Cluster and Grid ...
SemreX: Towards Large-Scale Literature Information Retrieval and Browsing with Semantic Association* Xiaomin Ning, Hai Jin, Hao Wu Cluster and Grid ...
热门文档