Surface Water Extraction from High-Resolution Remote Sensing Images Based on TA-UNet3+

Accurate extraction of surface water information from remote sensing images is crucial for water resources management, environmental monitoring and other fields. However, this task still faces a series of challenges due to factors such as the diversity of land cover, the intersection of water bodies and the surrounding environment, and complex occlusion of vegetation. In order to improve the accuracy of surface water extraction, a TA-UNet3+ network model suitable for high-resolution remote sensing images is proposed based on the optimization of U-Net3+ network. At the encoder end, the window attention embedding module is introduced layer by layer from the depth feature to the shallow layer, and the local attention from the deeper feature is gradually embedded into the lower-level feature to improve the semantic comprehension ability of the feature map. The TA-ASPP module, which combines threshold attention and depth separability, is introduced, which effectively improves the extraction efficiency of feature information. The multi-scale fusion module is modified on the decoder side, and the learnable dense upsampling convolution and deep separation convolution are used to replace the original bilinear interpolation and ordinary convolution, which significantly reduces the computational complexity while ensuring the accuracy. The dataset consists of parts of Chongqing city under different scenarios. Experimental results show that the TA-UNet3+ network model is better than the semantic segmentation network in terms of accuracy, recall, F1 and IoU, and shows higher surface water extraction accuracy. © 2025 Journal of Computer Engineering and Applications Beijing Co., Ltd.; Science Press. All rights reserved.