A New Automatic Antarctic Surface Crevasse Extraction Method Based on Deep Learning Using Sentinel-1 SAR Data

Over Antarctic ice shelves, the development of crevasses is abundant. They play important roles in ice shelf calving events, meltwater transportation, surface energy balance and many other dynamic processes. Information about the crevasses' distribution is critical for understanding the evolution of Antarctic ice sheet. Due to the complicated concrete background of ice shelf surface and the diversity of surface crevasses, detecting crevasses high precisely is still challenging. In this study, we proposed an improved D-LinkN et network to identify crevasses over Antarctic ice shelves based on Sentinel-1 SAR imagery. The improved network's encoder part consists of Res2Net modules with attention module SE blocks to achieve a fast focus on crevasses. The center part is composed of dilated convolutions, which increase the receptive field of feature points. The experimental results on Antarctic ice sheet show our method exhibits better performance in crevasse extraction compared with those methods based on U-net, FCN, SegNet and D-LinkNet.