GeoAI for Flood Inundation Mapping

What This Research Covers

• Systematic spatial comparison of OWP HAND-FIM and HEC-RAS hydraulic simulation outputs — characterizing agreement zones, inundation boundary errors, and conditions where HAND-FIM's simplified terrain assumptions diverge from physics-based baselines.
• Physics-aware surrogate ML model trained on paired HAND-FIM/HEC-RAS outputs across diverse flood events, watersheds, and flow regimes, learning to predict the spatial correction needed to bring HAND-FIM closer to hydraulic accuracy.
• Hybrid AI pipeline benchmarked at a 35% Critical Success Index improvement over HAND-FIM baselines and a 10,000× speed-up over full HEC-RAS simulation — unlocking real-time ensemble flood forecasting at national scale.
• Scalable architecture designed for NOAA OWP's national HAND-FIM framework, driven by NWM streamflow inputs and deployable across CONUS catchments without per-watershed retraining.
• Deep learning time-series models (LSTM) applied to weather and hydrologic forecasting, extending the GeoAI approach into climate and meteorological prediction contexts.

Why It Matters

NOAA's operational HAND-FIM system generates millions of flood predictions daily but its simplified terrain assumptions introduce systematic errors in complex floodplains. A surrogate model that bridges HAND-FIM's computational speed with HEC-RAS physical accuracy changes what is possible in real-time flood warning — ensemble probabilistic maps, scenario-based planning outputs, and high-resolution inundation extents become feasible where full hydraulic simulation was never an option. The 10,000× speed advantage is not an incremental improvement; it is the difference between a batch overnight run and a live operational product.