遥感反演中参数的不确定性与敏感性矩阵

Uncertainty and Sensitivity Matrix of Parametersin Inversion of Physical BRDF Model

多角度遥感中二向性反射 (BRDF)物理模型的反演 ,是当前国际遥感界学术研究的热点之一。反演中先验知识的表达和利用 ,参数敏感性的判别和处理 ,已提出了若干年 ,但只是到最近 ,才有了将参数的先验概率引入拟合误差平方加权 ,及定义参数对观测值敏感性的尝试。该文简述了这些最新进展 ,并首次建议了反演中参数的“不确定性与敏感性矩阵”的定义。并以作者最近的反演实例 ,说明这样定义的矩阵 。

PhysicalBRDFmodelsareusuallyverycomplexanddifficulttoinvert Weusuallyneedtoemploya prioriknowledgeinthisorthatway ,fixsomeparametervaluesandinvertsomeothers Usuallymostofusagree thatnon -sensitiveparametersshouldbefixed Buttherehasnotbeenanyconsen -susonhowtodefinethe sensitivityofaparameterininversion LiandStrahler ,LiandWangalsosuggestedthatonlythosethemostsen sitiveandmostuncertainparametersshouldbeinvertedbyusingasubsetofobservations Buttheyfailedtospell outhowtodeterminesuch“mostsensitiveandmostuncertain” parametersandhowtofindsuchasubsetofob servations ThislackingofconsensusandquantitativerulesmakesinversionofphysicalBRDFmodelsacase -by -case“trick”oran“artbutscience” WetriedtodevelopageneralframeworkforBRDFmodelinversion Itisbasedonaccumulationofknowl edgeandaninversionstrategywhichwecalledMulti-stage ,Sample -directionDependent,Target-decisions (MSDT) Presently ,ourknowledgeinclude :1)DTM ;2 ) previousland -coverclassification ;3)seasonalchange patternoftheseland -covers ;4 )rightmodelforeverytypeoflandcovers ;5) physicallimitations (ornone) ofeachparameterineachmodel;6)abestguessofeachparametervalueandtheuncertaintyofsuchguess . OurMSDTinversionstrategyisbasedonanUncertaintyandSensitivityMatrix (USM )ofparametersat givendirections/bandsofobservations Itsdefinitionissomehowanalogoustothepartialderivativematrixused inNewtonmethodsforminimization ,buttherearethreesignificantdifferences :suchguess 1)Theuncertainty oftheinitialguessistakenintoaccount ;2 )Itislessdependentontheinitialguess ;3)Allelementshavethe sameunitandthereforequantitativelycomparable AnexampleofUSMfromLi -StrahlerGOMSmodeland ASASsamplingwillbepresented ,anditisobviousfromthematrixwhatparametershouldbeinvertedfirst , andwhatsubsetofobservationsshouldbeused AnotherexampleofUSMfromSAILmodelandhemispherical samplingisalsopresented ComparisonbetweeninversionerrorsofusingdifferentsubsetsofsamplesshowUSM couldbeahelpfulconceptinBRDFinversion