A Stepwise Method for Change Detection in Large-Scale Polarimetric SAR Images

Abstract

In this paper, a stepwise method which consists of two main steps is proposed to tackle change detection in large-scale polarimetric Synthetic Aperture Radar (SAR) images. First, down-sample two registered polarimetric SAR images and calculate the corresponding Difference Image (DI), then spatial localization is conducted to orient sub-region which contains changed area in high probability. Second, polarimetric SAR sub-images are selected and they are trained together with down-sampled whole images by Convolutional Neural Network (CNN), where changed areas in sub-region are revealed and the next sub-region is generated accordingly. Repeat these two steps until all interested regions are detected. In general, it collects the whole but coarse information at first to locate important domains with changed areas and then analyzes them for accurate detection result and generate the next sub-region for further detection. Experiment results show that the proposed method performs well in detecting changed areas in large-scale polarimetric SAR images.