Water Surface Mapping from Sentinel-1 Imagery Based on Attention-UNet3+: A Case Study of Poyang Lake Region

Abstract

The mapping of water surfaces is important for water resource and flood monitoring. Synthetic Aperture Radar (SAR) images can be used to monitor water bodies and detect floods over large areas. To address the problem of low identification accuracy in different time phases and different scales of water area, a water surface mapping method based on Attention-UNet3+ with SAR images is proposed in this paper. In the model, full-scale skip connections are introduced for combining feature maps from different scales and improving the accuracy of narrow water identification; the spatial attention module is used to obtain the importance of each connected feature, which can reduce the number of false alarms caused by speckle noise and water shadows in SAR data; the deep supervision module is used to learn hierarchical representative features from comprehensive aggregated feature maps to provide the periodic output capability of the model and meet the needs of rapid and large-scale water identification. The effectiveness of Attention-UNet3+ is verified by experiments in the Poyang Lake region with Sentinel-1 SAR images. The results show that the proposed Attention-UNet3+ outperforms the conventional threshold segmentation and deep learning models such as UNet, Deepvlabv3+, and SegNet, with an average IOU/Kappa value of 0.9502/0.9698. Multitemporal Sentinel-1 images in 2021 covering Poyang Lake are used for time series water surface mapping with the proposed method, and it is found that the detected water area of Poyang Lake has a good correlation with the corresponding water level values at observation stations. The Pearson coefficients are about 0.96. The above results indicate that the proposed method achieves good water surface mapping performance.

Keywords

• Synthetic Aperture Radar (SAR)
• water surface mapping
• deep learning
• Attention-UNet3+
• semantic segmentation