Water Body Automated Extraction in Polarization SAR Images With Dense-Coordinate-Feature-Concatenate Network

Abstract

Synthetic aperture radar (SAR) water body extraction is of great significance for many applications, such as flood disaster monitoring, coastline change detection, and water resources management. However, the previous research works have mainly focused on single-polarization SAR images, and the dual-polarimetric information has not been comprehensively utilized, which result in the problem of incomplete water body extraction. In this article, we propose a model called dense-coordinate-feature-concatenate network to extract and fuse the water features of dual-polarimetric SAR images, overcome the influence of ground interferences, and improve the accuracy and generalization of the model. In this model, the improved dense block module is used to extract and transfer the water features in the dual-polarimetric data, a coordinate attention module is used to reduce the loss of water body boundary, and a feature concatenate module is proposed to avoid information redundancy. Experimental results show that our proposed model can not only overcome the influence of roads and hill shades and achieve 98.4%, 84.62%, and 91.67% water body extraction accuracy in overall accuracy, intersection over union, and F1, respectively, on Gaofen-3 SAR images but also be directly applied to the water body extraction of Sentinel-1 SAR images.

Keywords

• Deep learning methods
• dual-polarimetric
• synthetic aperture radar (SAR)
• water body extraction