Accurate identification and detection of disasters are important research directions for remote sensing application. A dual attention mechanism U-Net model is proposed to address the problem of water body segmentation in flood disasters. Considering the limitations of optical remote sensing due to weather factors in flood disasters, we used SAR images. Firstly, based on the characteristics of SAR images, spatial attention and channel attention are embedded into the U-Net network, enabling the final segmentation results to have more detailed and accurate segmentation content. Secondly, to further improve the model’s performance and avoid overfitting, we cropped the data based on the distribution characteristics of water bodies in SAR images and performed data augmentation. Ultimately, we processed and trained the model using pre-disaster data, achieving accurate segmentation of water body changes before, during, and after the disaster.