2026 CSDMS meeting-022
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
Enhancement of low-fidelity operational flood inundation mapping through surrogate modeling
Supath Dhital,
(He/Him),The University of Alabama Tuscaloosa Alabama, United States. sdhital@crimson.ua.edu
Sagy Cohen, The University of Alabama Tuscaloosa Alabama, United States. sagy.cohen@ua.edu
Parvaneh Nikrou, The University of Alabama Tuscaloosa Alabama, United States. pnikrou@crimson.ua.edu
Anupal Baruah, The University of Alabama Tuscaloosa Alabama, United States. abaruah@ua.edu
Yixian Chen, The University of Alabama Tuscaloosa Alabama, United States. ychen223@ua.edu
Dinuke Munasinghe, The University of Alabama Tuscaloosa Alabama, United States. dsmunasinghe@ua.edu
Dipsikha Devi, The University of Alabama Tuscaloosa Alabama, United States. ddevi@ua.edu
Floods cause hundreds of deaths and billions of dollars in damage worldwide each year. While reliable Flood Inundation Mapping (FIM) is crucial for flood forecasting and analysis, these systems face a key challenge: high-fidelity (HF) physics-based models (e.g., HEC-RAS) accurately simulate flood dynamics but are resource-intensive and slow, making them impractical for emergencies and large-scale applications. Conversely, low-fidelity (LF) terrain-based methods (e.g., NOAA's operational Height Above Nearest Drainage (HAND)- based conceptual FIM framework) are faster but often less accurate, which can mislead emergency response. This study aims to bridge that gap by developing a U-Net Convolutional Neural Network (CNN)-based hybrid AI surrogate model (SM) that combines conceptual LF-FIMs with topographical and hydrological forcings and HF-FIM simulations. The SM is trained using extensive HEC-RAS simulations across various elevations and landscapes, integrating the LF NOAA OWP operational HAND-based FIM. This combination enables the model to learn from both conceptual terrain-water interactions and deep learning's capability to model complex nonlinear relationships. Embedding this knowledge helps the model understand fundamental physics of land-water interactions, serving as an attention mechanism that improves its ability to generalize across space and time. Independent evaluation of SM derived FIM across multiple US locations shows a median 14% improvement in the Critical Success Index, with a 38% enhancement in event-scale inference. Analyzing impact based on building footprints shows even more promising results, with adaptive improvements in building exposure metrics compared to LF-FIM. In addition, the computation time relative to the HF model shows a significant speedup provided by the SM. Therefore, this method offers portability and interpretability, presenting significant advantages over purely data-driven models and increasing scientific credibility. Although integrating LF and HF simulations from different framework configurations was technically complex, the results are impactful and scalable. The SM is incorporated into NOAA's operational framework through the FIM as a Service (FIMserv) platform, streamlining flood inundation mapping for nationwide deployment. This includes both retrospective FIM improvements and coverage of the entire forecast range (short-, medium-, and long-term). This demonstrates its potential as a scalable, reliable, and efficient component of national flood forecasting systems.
