Hierarchical Interactive Attention Res-UNet for Inland Water Monitoring With Satellite-Based SAR Imagery

Abstract

The extraction and change detection of river-lake boundaries has a wide range of applications. Especially in flood monitoring, river-lake boundary monitoring systems can provide important information support for relevant departments in flood prevention and relief. Aiming at the problem that the existing research has low accuracy and poor boundary quality in complex terrain, we propose a hierarchical interactive attention Res-UNet. By introducing a cross-dimensional attention mechanism with hierarchical interactions, the model's ability to perceive key features is improved. In addition, we propose contour-aware mixed loss, which pays more attention to difficult samples and edge pixels. Based on Sentinel-1A satellite data, we produce a synthetic aperture radar data set of river and lake water bodies. Through comparative experiments, the effectiveness of this method in improving water body segmentation accuracy and boundary quality is proved. The final F1 score is 0.973 and HD95: Hausdorff distance (HD) with a 95% confidence level is 2.219. Finally, we apply our methodology to study the changes in the water body near Wuhan Sancha Harbor during the 2020 Yangtze River flood, which proves the effectiveness of the scheme.