Authors: Reihaneh Zarrabi – University of Alabama
Title: Improving Operational Flood Mapping Across the CONUS Using Machine Learning and Bathymetric Adjustment
Presentation Type: Lightening talk
Abstract: Accurate continental-scale flood inundation mapping is critical for forecasting, hazard mitigation, water resources management, and community resilience, but its reliability remains constrained by how river channels are represented. In large-scale applications, channel geometry is often estimated with simplified methods, and channel bathymetry is not explicitly represented, reducing the realism and accuracy of flood predictions. This study presents a two-stage data-driven framework to address these limitations. First, advanced tree-based machine learning models are developed using a preprocessed HYDRoacoustic dataset supporting Surface Water and Ocean Topography (HYDRoSWOT) and predictors from the National Hydrography Dataset Plus Version 2.1 (NHDPlusV2.1) to estimate bankfull and mean-flow channel width and depth. The models are then applied to generate a continental-scale channel geometry dataset covering approximately 2.7 million stream reaches across the conterminous United States (CONUS). In the second stage, this dataset is integrated into the Height Above Nearest Drainage Flood Inundation Mapping framework (HAND-FIM), NOAA’s operational flood forecasting system used by the Office of Water Prediction, through the Bathymetric Adjustment of Rating Curves and Stage Shift (BARC-SS) method, which updates synthetic rating curves using the machine learning-derived channel geometry and shifts the stage reference from the water surface to the channel bed. The machine learning models achieved R² values of 0.85 for channel width and 0.69 for channel depth, outperforming existing regional methods by 30% and 76%, respectively. Across six U.S. basins, BARC-SS improved flood extent and building-impact predictions compared to the original model, with the critical success index increasing by up to 113% and the true positive rate by more than 130% in basins where the original model underestimated flooding. Overall, this work provides a scalable pathway for improving operational flood inundation mapping and supporting better flood preparedness and response.