Improving the fidelity and performance of a conceptual flood inundation mapping approach using a machine learning-based surrogate model

This study focuses on enhancing the accuracy of Flood Inundation Mapping (FIM) by utilizing a surrogate modeling approach. The Height Above the Nearest Drainage (HAND) method is used as our baseline FIM approach. The terrain-based HAND-FIM framework was developed to allow large-scale applications at low computational costs. A Surrogate Model (SM) is constructed using machine learning-based methodologies to emulate the high-fidelity Hydrologic Engineering Center-River Analysis System (HEC-RAS) model. HAND-FIM, generated using streamflow data from the National Water Model, serves as the input to the SM, while the flood extent predicted by HEC-RAS for the same event serves as the target. Results demonstrate that SM reduces false alarms in HAND-FIM by 18% while improving the Critical Success Index score by 26%. Integrating the SM offers a promising approach for enhancing flood prediction accuracy, mitigating HAND-FIM limitations, and providing fast, cost-effective solutions for operational FIM applications, especially in data- and resource-limited regions. • A Random Forest-based surrogate model was developed to enhance the accuracy of conceptual FIM. • Effective cross-model knowledge transfer was achieved between conceptual and hydrodynamic FIM approaches. • Surrogate model significantly reduced the false alarm rate of the conceptual FIM approach. • The method enabled rapid, accurate flood mapping with lower computational cost. • It demonstrated promising potential for spatial generalization in data- and model-scarce areas.