Urban Area Analysis in Single-polarized SAR Images Based On Unsupervised Deep Learning

Abstract

Urban mapping from remote sensing images is important for monitoring urbanization. In this paper, we propose an unsupervised learning algorithm for high-resolution single-polarized synthetic aperture radar (SAR) image to extract man-made targets for urban area analysis. The proposed method mainly focuses on the special physical characteristics of man-made targets that are different from natural areas. Without polarimetric information, we propose the sub-band scattering pattern based on time-frequency analysis to describe the physical properties of targets, and then design an end-to-end neural network to learn the latent features and potential clusters. The proposed method is evaluated on three different urban areas acquired at C-band by Sentinel-1 and Gaofen-3, and X-band by TerraSAR-X, respectively. The experiments present the visualized result of man-made targets extraction and analyze some specific targets to show the effectiveness of our proposed method.