Progressive Residual Phase-Aware Token-Mixing Architecture for Global Sentinel-1 Flood Mapping

Synthetic Aperture Radar (SAR) enables reliable flood mapping under adverse weather and lighting conditions, but operational products remain affected by speckle noise, heterogeneous land cover, and confusion between low-backscatter surfaces and open water. We present Pr-ResWaveUNet, a lightweight deep-learning architecture for SAR-based flood segmentation that combines a residual encoder–decoder with progressive, phase-aware token-mixing wave blocks in skip connections. The model uses three-channel Sentinel-1 input: normalized pre-event VV, normalized post-event VV, and their radiometric difference (ΔVV), enhancing sensitivity to flood-induced changes while mitigating noise. Trained on the S1GFloods dataset (42 events, 5,360 image pairs), the model achieves 94% IoU, 97% F1-score, and 98% accuracy, outperforming U-Net baselines with only ~2.6M parameters. Tested on independent 2025 floods in France, Romania, and Pakistan, it reached 86–89% IoU. Integrated into the SaferSat toolbox within SaferPlaces, it enables automatic, global-scale SAR flood mapping.