Class-based kernels selection for albedo inversion by kernel-driven BRDF model

Kernels are always pre-determined in current kernel-driven model applications, but they seem to have some disadvantages in the requirement for more accurate remote sensing because one kernel combination is used for the inversion of all land cover types. In this paper, we use 28 different multi-angular data sets, which represent major types of land cover, to find the relations of different kernel selections with land cover types. The kernel combinations in the models we compare are volume kernels of Ross-Thick, Ross-Thin and geometric optical kernels of Li-Transit, Li-SparseR and Li-Dense. The airborne multi-angle TIR/VNIR image system (AMTIS) data set, which was obtained in Shunyi county of Beijing, China in April 2002, was used for the inversion. The inversion results of pre-determined kernel selections and class-based kernel selections are compared.