A Cross-Domain Semi-Supervised Water Body Extraction Framework for SAR Images

In recent years, with the intensification of global climate change, extreme weather events have become more frequent, leading to severe economic losses and social impacts worldwide due to flooding. Synthetic Aperture Radar (SAR) technology, as an all-weather, all-time remote sensing tool, can acquire high-quality Earth observation data under adverse weather conditions, providing unique advantages for real-time monitoring and assessment of flood disasters. Although SAR images can provide valuable information, differences in characteristics between satellite systems and variations in imaging across different regions present numerous challenges for model training, which in turn affect model performance. We propose a cross-domain semi-supervised water body extraction framework for SAR images, building upon the unimatch semi-supervised framework. The proposed framework introduces a cross-domain multiscale feature metric and an enhancement strategy based on SAR imaging characteristics, enhancing model generalization while reducing disparities between target and source domains. Experimental results demonstrate that the proposed method effectively enhances model robustness and performance across diverse datasets.