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

Synthetic Aperture Radar (SAR) provides a reliable basis for flood mapping underadverse weather and illumination conditions; however, operational products arestill affected by speckle noise, heterogeneous land cover, and confusion betweennaturally low-backscatter surfaces and open water. To address these limitations, wepresent Pr-ResWaveUNet, a lightweight deep-learning architecture for SAR-basedflood segmentation that integrates a residual encoder–decoder with progressive,phase-aware token-mixing wave blocks in the skip connections. The model uses athree-channel Sentinel-1 input comprising normalized pre-event VV backscatter,normalized post-event VV backscatter, and their radiometric difference (ΔVV),enhancing sensitivity to flood-induced changes while mitigating speckle and lowsignal-to-noise effects.Pr-ResWaveUNet was trained and evaluated on the S1GFloods dataset, covering 42global flood events from 2016 to 2022 with 5,360 Sentinel-1 GRD image pairs andreference labels. The model achieves an intersection- over-union (IoU) of 94%, anF1-score of 97%, recall of 97%, precision of 97%, and an overall accuracy of 98%,outperforming standard U-Net baselines and matching representative Transformerbasedmodels. It is computationally efficient, with ~2.6 million parameters, aboutan order of magnitude fewer than comparable models.Generalization was assessed on three independent flood events in 2025 in France,Romania, and Pakistan using reference maps derived from Sentinel-1 and Sentinel-2imagery, digital elevation models (DEMs), and land-cover data. Across theseevents, the model achieved IoU values of 86–89% and F1-scores of 93–94%, producing spatially coherent flood maps with reduced speckle fragmentation andpreserved connectivity along narrow channels. These results demonstrate that Pr-ResWaveUNet provides an effective, efficient, and scalable solution for large-scaleSAR flood mapping.For operational deployment, the model is integrated into the SaferSat toolbox withinthe SaferPlaces platform, enabling fully automatic, global-scale flood mapping.