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。

Physical BRDF models are usually very complex and difficult to invert. We usually need to employ a priori knowledge in this or that way, fix some parameter values and invert some others. Usually most of us agree that non-sensitive parameters should be fixed. But there has not been any consensus on how to define the sensitivity of a parameter in inversion. Li and Strahler, Li and Wang also suggested that only those the most sensitive and most uncertain parameters should be inverted by using a subset of observations. But they failed to spell out how to determine such"most sensitive and most uncertain" parameters and how to find such a subset of observations. This lacking of consensus and quantitative rules makes inversion of physical BRDF models a case-by-case"trick"or an"art but science" . We tried to develop a general framework for BRDF model inversion. It is based on accumulation of knowledge and aninversion strategy which we called Multi-stage, Sample-direction Dependent, Target-decisions (MSDT) Presently, our knowledge include :1)DTM;2 ) previous land-cover classification;3)seasonal change pattern of these land-covers;4)right model for every type of landcovers;5) physical limitations (or none) of each parameter in each model;6)a best guess of each parameter value and the uncertainty of such guess. Our MSDT inversion strategy is based on an Uncertainty and Sensitivity Matrix (USM )of parameters at given directions/bands of observations. Its definition is somehow analogous to the partial derivative matrix used in Newton methods for minimization, but there are three significant differences: 1)The uncertainty of the initial guess is taken into account; 2)It is less dependent on the initial guess;3)All elements have the same unit and therefore quantitatively comparable. An example of USM from Li-Strahler GOMS model and ASAS sampling will be presented, and it is obvious from the matrix what parameter should be inverted first, and what subset of observations should be used. Another example of USM from SAIL model and hemispherical sampling is also presented Comparison between inversion errors of using different subsets of samples show USM could be a helpful concept in BRDF inversion.