SE-UNet: Channel Attention Based UNet for Water Body Segmentation from SAR Image

Water body extraction based on SAR remote sensing images plays an important role in water environment analysis and dealing with flood disasters. In recent years, a lot of algorithms based on encoding-decoding network models have been proposed in the field of water body segmentation. However, these algorithms have fewer considerations of channel attention information in the decoding process, which brings disadvantages for achieving higher accuracy of water body segmentation. In this paper, a channel attention based UNet network was proposed to improve discrimination of water body regions by adding Squeeze and Excitation block in the decoding process. The experimental results show that the proposed method achieves 85.17%, 90.79% and 94.22% of Iou, Recall and Precision performance for water segmentation on our SAR dataset, respectively. Compared with the UNet model, the three evaluation metrics are improved by 5.15%, 5.22%, and 8.65%, respectively. The proposed SE-UNet also demonstrated better performance compared to SegNet, PSPNet in SAR water segmentation.