SaferSat: The Saferplaces’s Operational Sentinel-1 Toolbox for Multi-Temporal Flood Extent Mapping, Water-Depth Estimation and Impact Assessment

Operational flood intelligence for emergency response and insurance, providing a rapid overview of impacted land, population, and economic damages, requires mapping solutions that remain reliable under adverse observational conditions and across diverse landscapes. Although Sentinel-1 SAR provides consistent global, all-weather and day-and-night coverage, automated flood extraction is challenged by speckle noise, land-cover heterogeneity, and confusion between floodwater and permanent low-backscatter surfaces. These limitations highlight the need for approaches that exploit temporal backscatter changes while maintaining global robustness and computational efficiency.

We present SaferSat, a fully automated Sentinel-1 toolbox for flood-extent mapping, water-depth estimation, and impact assessment. SaferSat is part of SaferPlaces (saferplaces.co), a global Digital Twin platform for flood risk intelligence supporting emergency response and insurance applications. Central to the framework is Pr-RWU-Net (Progressive Residual Wave U-Net), a lightweight deep-learning model with 2.6 million trainable parameters, designed to detect flood-induced backscatter changes using VV-polarized SAR imagery. The model uses a three-channel input; pre-event VV, post-event VV, and their radiometric difference, enhancing inundation sensitivity while mitigating VH instability for global deployment.

SaferSat provides end-to-end processing: automated data retrieval, multi-date flood inference, and Maximum Flood Extent generation. To reduce SAR ambiguities, it generates auxiliary layers: a vegetation mask for SAR "blind spots" and a low-backscatter anomaly mask for permanent dark features. Flood extent layers are integrated with the FLEXTH model and GLO-30 or local high-resolution LiDAR DTMs for water-depth reconstruction. The system also analyzes acquisition patterns to predict short-term revisit opportunities. Impact assessment intersects flood extents with JRC GHS-POP and ESA WorldCover datasets.

The Pr-RWU-Net model was trained on the S1GFloods dataset, containing 5,360 paired pre- and post-event Sentinel-1 GRD images across 42 flood events from 2016–2022. Binary flood masks were generated via semi-automated thresholding and expert quality control. Evaluation on the test split achieved an IoU of 90.0%, F1-score 94.6%, Recall 95.6%, Precision 93.8%, and overall accuracy 96.6%.

Operational applicability was demonstrated on three 2025 flood events: Romania, Pakistan, and France. SaferSat flood extents closely matched SAR manual driven flood references (IoU 89–92%) and CEMS products (IoU 85–88%). Water-depth estimation against a reference hydrodynamic model yielded a MAE of 34–40 cm and correlation R of 0.78–0.82. For a 260 km² flood in Romania, the full processing chain completed in ~3 minutes on a standard CPU, demonstrating suitability for rapid, large-scale deployment.

SaferSat is available globally through SaferPlaces, supporting emergency response and insurance applications. Future developments aim to enhance SaferSat globally via integration of commercial satellite data to reduce revisit time and rapid hydrodynamic modeling to address radar limitations.