温度反演经典文章(14)

Author's personal copy

26

Z.-L.Lietal./RemoteSensingofEnvironment131(2013)14–37

3.2.3.SimultaneousretrievalofLST,LSEs,andatmosphericpro les

AlthoughthesimultaneousLSTandLSEretrievalmethodsreviewedabovecanaccuratelyobtaintheLSTandLSEsiftheatmo-sphericcorrectionsareperformedproperly,accurateatmosphericpro lesareusuallyunavailablesynchronouslywithTIRmeasure-ments,andthustheaccuracyoftheretrievedLSTandLSEscanbede-graded.AnidealsolutionistosimultaneouslyretrievetheLST,LSEs,andatmosphericparameters(e.g.,atmosphericpro les)(Maetal.,2002).BecausethenarrowbandwidthofferedbyhyperspectralTIRsensorswiththousandsofchannelscanimprovetheverticalresolu-tionandallowatmosphericpro lesandsurfaceparameters(LSTandLSEs)tobeobtainedmoreaccurately(Chahineetal.,2001),severalmethodshavebeenproposedtoretrievesimultaneouslythesurfaceandatmosphericparameters.Therepresentativesofthesemethodsarethearti cialneuralnetwork(ANN)method(Wangetal.,2010)andthetwo-stepphysicalretrievalmethod(Maetal.,2002,2000).

3.2.3.1.Arti cialneuralnetwork(ANN)method.BecauseanANNcanrobustlyperformhighlycomplex,non-linear,parallelcomputations,ANNshavebecomeincreasinglyutilizedbytheremotesensingcom-munity(Mas&Flores,2008).ANNsresemblethebrainintwoas-pects:theyacquireknowledgethroughalearningprocess,andstoretheacquiredknowledgeusinginterneuronconnectionstrengths(Mas&Flores,2008).Therefore,ANNsrepresentmassivelyparalleldistributedprocessorsthatcanacquireexperientialknowledgeandmakethatknowledgeavailableforuse.

ThemainadvantagesofANNmethodsoverconventionalretrievalmethodsaretheirabilitytolearncomplexpatterns,generalizationtonoisyenvironments,andincorporationofbothknowledgeandphys-icalconstraints(Mas&Flores,2008).BecauseofANNs'powerfulnon-linearretrievalabilities,anumberofattemptshavebeenmadetodevelopneuralnetworkstoretrieveboththesurfaceandatmo-sphericbiophysicalvariableswithoutexactknowledgeofthecom-plexphysicsmechanisms.Forexample,Maoetal.(2008)usedanANNtoestimatetheLSTandLSE,whileAiresetal.(2002b)andBlackwell(2005)usedanANNtoretrieveatmosphericpro les.Tore-ducetheeffectofcouplingbetweenthesurfaceandatmosphereontheretrievalaccuracy,Airesetal.(2002a)proposedusinganANNtoretrieveboththeatmosphericandsurfacetemperatures,andWangetal.(2010)attemptedtoestablishaneuralnetworktosimul-taneouslyretrievetheLST,LSE,andatmosphericpro lesfromhyperspectralTIRdata.ThepreliminaryresultsdemonstratedthatANNscanbeusedtosimultaneouslyretrievetheLST,LSEsandatmo-sphericpro lesfromhyperspectralTIRdatawithacceptableaccuracyforsomeapplications.RMSEsofLSTandtemperaturepro lesintroposphereareabout1.6Kand2K,respectively;RMSEofWVisaround0.3g/cm2.RMSEofLSEislessthan0.01inthespectralinter-valfrom10μmto14μm(Wangetal.2013).

However,becauseANNsperformlikeblackboxesandcanproducecorrespondingoutputsfromanygiveninputs,theretrievalprocesscannotbewellcontrolled,anditisdif culttointerprettheweightsassignedtoeachinputandimprovetheoutputduetothecomplexnatureofthenetwork.Inaddition,theimplementationofanANNde-pendslargelyonitsarchitectureandthetrainingdata(Mas&Flores,2008).Itisdif culttodeterminethearchitecturesandlearningschemesforanANN,whicharedirectlyrelatedtoitsabilitytolearnandgeneralize.Althoughoneortwohiddenlayersarerecognizedtobeenoughformostproblems(Airesetal.,2002b;Mas&Flores,2008;Sontag,1992),anumberofexperimentsarestillrequiredtode-terminewhatarchitecture-relatedparameterswillimprovetheaccu-racy,suchasthenumberofinputandhiddennodes,theinitialweightrange,theactivationfunctions,thelearningrateandmomentum,andthestoppingcriterion.Untilnow,noANNarchitectureisuniver-sallyacceptedforaparticularproblem.Thecharacteristicsofthetrainingdata,suchasthesizeandtherepresentativeness,arealsoof

considerableimportance.Theuseoftoofeworunrepresentativetrainingsampleswillresultinanetworkthatcannotaccuratelyre-trievetheoutputs,whiletheuseoftoomanytrainingsamplesre-quiresmoretimeforlearning.Becausephysicalunderstandingisnotrequired,ANNmethodsmayberegardedasempiricalmethods.However,theirresultscanbeusedtoprovideinitialguessesforfur-therimprovementsinthephysicalretrievalmethods(Motteleretal.,1995).MoredetailedinformationabouttheapplicationofANNscanbefoundintheworkofMasandFlores(2008).

3.2.3.2.Two-stepphysicalretrievalmethod(TSRM).Becausethemea-suredradianceattheTOAisafunctionofthesurfaceandatmosphericparameters,thesurfaceandatmosphericvariablescantheoreticallybeobtainedbyselectingappropriatechannelsevenfrommultispec-traldata.Maetal.(2000)madeaninitialattemptatsimultaneouslyretrievingtheLSTandatmosphericpro lesbyassumingthattheLSEisinvariantwithintheMIRchannelsandalsoinvariantwithintheTIRchannelsandbyignoringthesolarcontributioninMIRchan-nels.However,theseroughassumptionsmayleadtodegradedaccu-raciesinthetroposphere.Alongthislineofreasoning,Maetal.(2002)furtherconsideredthesolarcontributionandproposedanex-tendedtwo-stepphysicalretrievalmethodthatsimultaneouslyex-tractstheLST,theLSE,andtheatmosphericpro lesfromMODISdata.

ThemainideaunderlyingtheTSRMinheritsthatofatmosphericpro leretrieval.The rststepistotangent-linearizetheatmosphericRTEwithrespecttotheatmospherictemperature-humiditypro les,theLST,andtheLSEs.Giveninitialguessesforthoseatmosphericandsurfacevariables,asetofequationsbasedonthetangent-linearizedRTEcanbederivedusingtheremotelysensedmeasurements(Lietal.,1994;Maetal.,1999;Smith,1972).Atthesametime,theprinciple-component-analysis(PCA)techniqueandtheTikhonovregularizationmethodareemployedtoreducethenumberofunknownsandstabilizetheill-posedproblem(Maetal.,2000;Smith&Woolf,1976),whichmakesthesolutionoftheseequationsstableanddeterministic.AccordingtothestatisticalanalysisintheworkofMaetal.(2000,2002),only vetemperatureandthreewatervaporeigenvectorscanexplainalloftheinformationof40atmospherictemperatureandwatervaporlevels,respectively.Inthesecondstep,theNewtonianiter-ationalgorithmisutilizedwiththeregularizedsolutionastheinitialguesstoobtainthe nalmaximumlikelihoodsolutionoftheatmo-spherictemperature-humiditypro le …… 此处隐藏：5177字，全部文档内容请下载后查看。喜欢就下载吧 ……