SAGE-UNet: A mixture-of-experts spatially-aware gated enhancement U-Net for flood water segmentation from Sentinel-1 SAR and post-event impact assessment—learning from flood events in Germany and Spain

The increasing frequency and intensity of floods pose substantial risks to residential areas, infrastructure, and economic systems, highlighting the need for reliable flood mapping to support emergency response and damage assessment. Even though synthetic aperture radar (SAR) facilitates all-weather flood monitoring, speckle effects and intricate surface backscattering mechanisms make precise flood delineation challenging. To address these, we propose a hybrid flood-mapping framework centered on a Mixture-of-Experts Spatially-Aware Gated Enhancement U-Net (SAGE-UNet) for post-event segmentation using Sentinel-1 SAR imagery. This architecture incorporates polarization-specific denoising to minimize the impact of speckle effects and a spatially aware bottleneck to improve the representation of contextual features. To further reduce false alarms and missed detections, particularly in vegetated regions, we introduce a lightweight bi-temporal refinement module that combines discrete wavelet transform (DWT)-based change detection with context-guided morphological reconstruction using normalized difference vegetation index (NDVI)-derived vegetation constraints and hazard-related spatial priors. The framework was evaluated on two major European flood events (Hannover, Germany, December 2023; Valencia, Spain, October 2024). The experimental results demonstrated that the standalone SAGE-UNet achieved IoU improvements of 1.7% and 2.3% over the strongest CNN- and transformer-based baselines on the Hannover and Valencia datasets, respectively. After bi-temporal refinement, the complete framework achieved relative IoU improvements of 7.8% and 15.8%, over the strongest competing methods on the Hannover and Valencia datasets, respectively. The refinement also improved spatial consistency by reducing false alarms and missed detections and better delineating flooding beneath vegetation. LULC analysis also confirmed the framework’s utility for flood exposure assessment, identifying agricultural areas and infrastructure within floodplains as the most affected classes. These results demonstrate the framework’s potential for operational SAR-based flood mapping and impact assessment.