堆栈自编码器降维的内陆水体叶绿素a反演模型及应用
Abstract
TheconcentrationofChlorophylla(Chl-a)isamainindexofwatereutrophication,and
itisoneoftheimportantfactorsaffectingthespectralcharacteristicsofwaterreflectance.Due
totheinfluenceofterrigenousmaterials,theopticalcharacteristicsofinlandwaterbodiesare
complicated,whichbringsgreatdifficultiestotheretrievalofChl-aconcentration.Withthe
rapiddevelopmentofhyperspectralandmachinelearningtechnologies,itprovidesatechnical
solutionforquantitativeinversionofChl-aconcentrationininlandwater.
Inthispaper,Baiyanglakeisselectedastheresearcharea,andthemeasured
hyperspectraldataofwaterbodyandremotesensingdataofZY1F_AHSIhyperspectral
satelliteareusedtosolvetheproblemsofinsufficientfeatureextractionabilityoflinear
dimensionalityreductionmethodandlowlearningefficiencyandpoorgeneralizationability
ofneuralnetworkmodelintheprocessofcombininghyperspectraldataandmachinelearning
algorithm.AstackautoencoderparticleswarmoptimizationBPneuralnetwork
(SAE-PSO-BP)modelisproposedandappliedtomeasuredhyperspectraldataandsatellite
hyperspectraldata.Theresearchcontentsandconclusionsofthispapermainlyincludethe
followingaspects:
(1)ThispaperproposesaSAE-PSO-BPmodel,whichusesSAEnetworktoachievedata
dimensionalityreductionwhilepreservingwaterradiationinformationintheoriginalspectral
datatothegreatestextent,extractsthedepthcharacteristicsofthemeasuredwaterspectrum,
takestheinitialweightofBPneuralnetworkasthepositionvectorofparticles,andsearches
theoptimalvalueoftheinitialweightofthenetworkthroughparticleswarmoptimization
algorithm.Itreducestheprobabilityoflocalextremevalue,improvesthestabilityofmodel
andtheaccuracyofinversion,andexplainstheprincipleandrealizesthealgorithm