Fractional Impervious Surface Mapping on Multispectral Images With Visible Shadows via a Bundle-Based Sparse Unmixing Model

Impervious surface abundance (ISA) is an important indicator for monitoring urbanization and environmental disaster management processes. Commonly used spectral unmixing techniques extract ISA in the form of mixed pixels; however, in previous multispectral unmixing studies based on ISA estimation, scholars have seldom simultaneously considered urban shading and endmember variability, which has severely affected the accuracy of the results. To solve this problem, we propose a combination of endmember bundle extraction based on superpixel segmentation and sparse unmixing for the ISA estimation of multispectral images containing visible shadows (EESS), considering the city district of Taiyuan as an example. The results show that the endmember bundles extracted via EESS are comparable with the manually identified endmember bundles in terms of accuracy, and the error differences in the unmixing results do not exceed 3%. In addition, the sparse unmixing algorithm combined with the endmember bundle extraction technique can effectively distinguish impervious surfaces, vegetation, and soils, and the root-mean-square error, mean absolute error, and systematic error of the whole area are much lower than those of the traditional spectral unmixing method, which are approximately 5.52%, 5.86%, and 8.24% lower than those of FCLSU_MANU, respectively. Further analyses reveal that adding two types of endmembers, namely, urban shadows and mountain shadows, to the urban component model can effectively mitigate the impact of shadows on ISA estimation. Furthermore, the consistency of the EESS method across Sentinel and Landsat data demonstrates its strong generalizability, enabling it to adapt to the characteristics of different sensors.