Building Extraction From PolSAR Image Based on Deep CNN with Polarimetric Features

For polarimetric synthetic aperture radar (PolSAR) images, building extraction has been a challenging topic for long time in applications of land-use and land-cover analysis. Due to similar structures of buildings and such vegetation as forest, they often exhibit similar PolSAR scattering characteristics that are often difficult to distinguishing. Recently, deep Convolutional Neural Network (CNN) has been widely investigated for image processing with many promising results. This paper proposes a method that combines polarimetric features with the CNN network to realize the comprehensive utilization of polarimetric and contextual information of PolSAR data for the extraction of building areas in PolSAR images. Comparison experiments on both ESAR and EMISAR L-band PolSAR datasets show that the proposed method can generate better results for building extraction.