Abstract: Currently, the application of deep learning in the extraction of water bodies from high-resolution remote sensing images has become a research hotspot in the remote sensing field. Among them, algorithms based on the U-Net network have demonstrated good performance in water body extraction. However, there is scarce research that provides in-depth and detailed comparisons of the performance differences of different U-Net network algorithms in water body extraction tasks. Therefore, this article selects three convolutional neural networks, named U-Net, U-Net++, and Attention-U-Net, and based on the GID dataset, draws conclusions through experiments and quantitative analysis. The results indicate that: U-Net++ achieves the highest training accuracy, followed by U-Net and Attention-U-Net, with accuracies of 0.912, 0.907, and 0.899 respectively. U-Net++ exhibits superior edge extraction capability compared to the other two networks. In segmenting different types of water bodies and distinguishing non-water areas similar to water bodies in remote sensing images, U-Net++ shows significantly better extraction results, while U-Net and Attention-U-Net are prone to omission errors and exhibit suboptimal performance.

Key words: water body extraction, high-resolution remote sensing imagery, U-Net