Novel deep learning based SAR image processing

Abstract

In recent years, deep learning techniques have been widely used. However, in the study of synthetic aperture radar (SAR) terrestrial water detection, it is still a difficult task to support the training of deep network models due to the challenge of data acquisition and small sample size. This dissertation constructs and tests a SAR terrestrial water detection dataset Lake-SAR, which contains 30 scenes of Sentinel-1 SAR images covering 15 lakes such as Qinghai Lake and Poyang Lake, involving 9 provinces in China, with types including wind free area water and low wind area water. Meanwhile, this dissertation conducted experiments using classical deep learning image segmentation algorithms, among which the U-Net network has the best performance with an overall accuracy of 90.3%. The experimental comparative analysis forms the index benchmark, which can facilitate other scholars to further develop SAR land water detection related research on the basis of this dataset.