Physics-guided AI hybrid (UNet + FNO) achieves 0.90 F1 for flood mapping

Accurate flood mapping is critical for disaster response, but data-driven models often ignore the physics of water flow. A new paper from Gebre, Talreja, and Hashemi-Beni (accepted at IEEE RadarConf 2026) introduces a hybrid deep learning framework that embeds hydrodynamic principles directly into the training process. The model combines a UNet for fine-scale spatial details with a Fourier Neural Operator (FNO) to capture basin-scale hydraulic interactions, using multi-modal inputs from Sentinel-1 SAR, Sentinel-2 optical imagery, and digital elevation models. Physics-informed residual losses based on depth-averaged shallow water equations enforce mass and momentum conservation, ensuring predictions are physically coherent even in heterogeneous terrain.

The results across diverse floodplain settings are striking: the hybrid model achieves an Intersection over Union of 0.82 and an F1 score of 0.90 for flood extent mapping—significantly outperforming UNet-only and FNO-only baselines. Water depth is predicted with an RMSE of just 0.21m, and flow velocity with 0.15 m/s. Mass imbalance remains below 2.1%, confirming that the physics constraints are working. Ablation studies show that removing these constraints drastically degrades stability and accuracy. This work demonstrates that embedding hydrodynamic principles into deep learning yields more reliable, scalable flood predictions, offering strong potential for operational monitoring and real-time disaster management.