Flood Area Segmentation with Alos-2/Palsar-2 Intensity and Interferometric Coherence with Deep Learning

Modern deep learning techniques have the potential to accurately predict the extent of flooding over a wide area in near real-time from SAR satellite image. In this paper, we use U-Net and UNet++ to detect flood area from a pair of PALSAR-2 images acquired in the same area at two different stages of flooding and non-flooding. We propose adding interferometric coherence to SAR intensity for the training set of deep learning network. The results show improved accuracy and distribution of flood mapping compared with those using only intensity images.