Optimizing Subpixel Impervious Surface Area Mapping Through Adaptive Integration of Spectral, Phenological, and Spatial Features

Reliable subpixel impervious surface area (ISA) mapping at medium resolution is essential but difficult due to the complexity of land cover patterns within the urban/periurban area. In this letter, we proposed a framework to optimize subpixel ISA mapping performance with adaptive integration of features from spectral, phenological, and spatial dimensions. We utilized the recursive feature elimination to build the most discriminative feature pool. Then, the random forest (RF) model was adopted for the subpixel ISA mapping and the feature contribution quantification. We applied the proposed framework in two typical study sites and tested its utility by comparing it with three other subpixel mapping approaches. The experimental results suggested that the inclusion of complementary feature inputs beyond the spectral profile was beneficial in both study sites for identifying fractional imperviousness. In particular, the improvement was most pronounced for pixels suffering spectral variability or intra-annual land cover change. With the quantification of feature contribution to the RF model, we further illustrated the critical impact of environmental conditions on the feature adoption.