Water area extraction method based on Res2Net-Unet

Regarding the issues that traditional water area extraction algorithms for remote sensing images require preset thresholds, are vulnerable to interference from mountain or vegetation shadows, and have poor robustness at different spatiotemporal scales, a model based on the Res2Net-UNet is proposed to extract water area from remote sensing images. The model adopts the UNet's encoder-decoder architecture, deepens the model's level with the ResNet network, and performs multi-scale feature extraction and fine-grained feature expression by combining Res2Net. Moreover, the model integrates an attention mechanism to increase the weight of key features and fully utilize the spatial and channel information in the images. After that, the model adds a skip connection to better fuse the encoder and decoder to preserve the detailed features of the image. Experimental findings on the GF-1 dataset demonstrate that the Res2Net-UNet model significantly enhances the accuracy of water extraction in remote sensing images.