Robust Water Region Detection in SAR imagery using Cross-Domain Adaptation

Accurate identification of water regions in Synthetic Aperture Radar (SAR) imagery is essential for effective water resource monitoring and disaster management. However, this task remains challenging due to the significant variations in imaging characteristics from different satellite sensors and environmental conditions, which hinder model generalization. This study introduces a novel domain adaptation framework for extracting water areas from SAR images. Our framework adapts the semi-supervised FixMatch method, which leverages both labeled and unlabeled data in single domain tasks, for the cross-domain segmentation challenge. A cross-domain training strategy is introduced to enhance model robustness. Experimental results indicate that our method achieves a +7% improvement in IoU performance over conventional FixMatch semi-supervised learning and +10% gain compared to supervised learning methods.