A Comparative Domain Generalization Study for SAR Image-Based Flood Segmentation

Floods are among the most devastating natural disasters, causing severe human and economic losses. Their occurrence frequency has been increasing progressively. Hence, effective and reliable flood analysis is essential for mitigating catastrophic losses. In this regard, Synthetic Aperture Radar (SAR) satellites are invaluable tools for providing large-scale images aimed at flood mapping under all weather conditions. Methods specifically designed for SAR image-based flood mapping are commonly developed under the assumption that the training (source) and test (target) data are sampled from the same distribution. However, in many real-world scenarios, these distributions often differ due to factors such as geographic location and incident angle depending on the satellite, leading to distribution shifts (a.k.a. domain shift), which ultimately degrades model performance. In this study, we investigate domain generalization approaches in combination with segmentation networks for the purpose of SAR based flood mapping. During the model’s training, each flood event is treated as a distinct source domain, with the objective of minimizing the domain shift among them to obtain a more robust model for unseen flooding events. Experiments conducted on the Sen1Floods11 dataset demonstrates an improvement in segmentation performance, with domain generalization approaches 3% in terms of IoU and F1 scores.