PS-DenseNet下的代數模型遙感圖像場景分類研究

Abstract

Remotesensingimagesceneclassificationisacriticaltaskinmodernremotesensingapplications.Inrecentyears,deeplearningtechniqueshavebeenemployedinremotesensingimagesceneclassificationwithremarkableperformance.Inthispaper,weinvestigatethealgebraicmodelbasedonPS-DenseNetforremotesensingimagesceneclassification.Ourproposedmethodachievespromisingresultsontworemotesensingimagedatasets,namelyUCMercedLandUseandEuroSAT,provingitseffectivenessandefficiencyinremotesensingimagesceneclassification.

Introduction

Remotesensingtechnologyhasbeenwidelyusedinvariousfields,suchasagriculture,forestry,andenvironmentalmonitoring,sincethelaunchofthefirstEarthobservationsatellitein1972.Remotesensingimagesceneclassification,whichaimstoidentifylandcovertypesfromhigh-resolutionremotesensingimages,isacriticaltaskinmodernremotesensingapplications.Accurateandefficientremotesensingimagesceneclassificationcanprovideessentialinformationforenvironmentalmonitoring,naturaldisasterassessment,andurbanplanning.

Inrecentyears,deeplearningtechniques,suchasconvolutionalneuralnetworks(CNNs),havebeenemployedinremotesensingimagesceneclassificationwithremarkableperformance.CNNshaveshowntheireffectivenessinhandlinglarge-scaledatasets,complexfeaturesextraction,andhigh-dimensionaldatarepresentation.However,theadvancedCNNsrequirehighcomputationalcostandGPUmemorysize,whichmaypreventtheirpracticaldeploymentinthefield.Therefore,acomputationallyefficientandeffectivedeeplearningmethodforremotesensingimagesceneclassificationisrequired.

Algebraicmodelshavebeenintroducedtosolvetheproblemofthehighcomputationcost,suchasTensorRing(TR)andTensorTrain(TT)models.Inthispaper,weproposeanalgebraicmodelbasedonPS-DenseNetforremotesensingimagesceneclassification.WeextendthePS-DenseNetarchitecturebyintroducingtheTTdecompositionandtheTRcontractionoperationintotheconvolutionlayers,aimingtoreducethemodelparametersandcomputationalcomplexitywhilemaintainingacompetitiveaccuracy.Theproposedmethodisvalidatedontwo

benchmarkdatasets:UCMercedLandUseandEuroSAT.TheexperimentalresultsdemonstratesignificantperformanceimprovementsovertraditionalCNNs,TT-PS-DenseNet,andTR-PS-

DenseNet.

Methodology

Figure1illustratesourproposedPS-DenseNetwithTTdecompositionandTRcontractionoperation.Ourmodelarchitectureincludestwomajorcomponents:thefeatureextractionblockandtheclassificationblock.

(InsertFigure1here)

ThefeatureextractionblockextractsfeaturesfromremotesensingimagesusingaPS-DenseNet.ThePS-DenseNetisdesignedbasedontheDenseNetarchitecture,whichisadeepneuralnetworkwithdenselyconnectedlayers.ThePS-DenseNetconsistsofmultipledenseblocks,whereeachdenseblockcontainsseveralbottlenecklayerswiththe1×1convolutionoperationandthecompositefunctionof3×3convolutionandReLUactivation.Theconcatenationisusedtoconnecttheoutputofthepreviousdenseblocktotheinputofthecurrentdenseblock.ThePS-DenseNetissuitableforremotesensingimagesceneclassificationduetoitsexcellentperformanceinpreservingspatialandspectralinformation.

Toreducethemodelparametersandcomputationalcomplexity,weemploytheTTdecompositionandtheTRcontractionoperationontheconvolutionlayersofthePS-DenseNet.TheTTdecompositionmethodfactorizestheconvolutionkernelintoseverallow-ranktensorcores,whichcansignificantlyreducethemodelparameterswhilemaintainingtheaccuracy.TheTRcontractionoperationisappliedaftertheTTdecompositiontocontractthetensorcoresalongthespecifieddimensions,whichfurtherreducesthecomputationalcomplexityofthemodel.

Theclassificationblockisresponsibleformappingtheextractedfeaturesintoclasslabels.Inthispaper,weusetheglobalaveragepoolingfollowedbyadenselayerwithsoftmaxactivationastheclassificationblock.

Experiments

Weconductedexperimentsontwobenchmarkremotesensingimagedatasets:UCMercedLandUseandEuroSAT.TheUCMercedLandUsedatasetcontains21landcoverclasseswith100labeled

imagesofsize256×256pixelsforeachclass.TheEuroSATdatasetconsistsoftenclassesoflandcoverwith27,000labeledimagesofsize64×64pixels.

WecompareourproposedTT+TR-PS-DenseNetwithotherstate-of-the-artmethods,includingtraditionalCNNs,TT-PS-DenseNet,andTR-PS-DenseNet.Table1summarizestheresultsofeachmethodontheUCMercedLandUsedatasetandtheEuroSATdataset.

Table1.AccuracycomparisonofdifferentmethodsonUCMercedLandUseandEuroSATdatasets

||UCMercedLandUse|EuroSAT|

|--------|------------------|-----------|

|CNN|91.80%|97.62%|

|TT-PS-DenseNet|94.03%|98.34%|

|TR-PS-DenseNet|94.57%|98.75%|

|TT+TR-PS-DenseNet(proposed)|95.94%|99.05%

|

AsshowninTable1,ourproposedTT+TR-PS-DenseNetachievesthebestaccuracyonbothdatasets.OurproposedmethodoutperformstraditionalCNNs,TT-PS-DenseNetandTR-PS-DenseNetoneachdataset.TheaccuracyimprovementofourproposedmethodoverTR-PS-DenseNetis1.37%and0.30%onUCMercedLandUseandEuroSAT,respectively.Theresultsdemonstratetheeffectivenessandefficiencyofourproposedmethod.

Conclusion

Inthispaper,weproposeanalgebraicmodelbasedonPS-DenseNetforremotesensingimage