FloodGenome: Interpretable Machine Learning for Decoding Features Shaping Property Flood Risk Predisposition in Cities

Understanding the fundamental characteristics that shape the inherent flood risk disposition of urban areas is critical for integrated urban design strategies for flood risk reduction. Flood risk disposition specifies an inherent and event-independent magnitude of property flood risk and measures the extent to which urban areas are susceptible to property damage if exposed to a weather hazard. This study presents FloodGenome as an interpretable machine learning model for evaluation of the extent to which various hydrological, topographic, and built-environment features and their interactions shape flood risk disposition in urban areas. Using flood damage claims data from the U.S. National Flood Insurance Program covering the period 2003 through 2023 across four metropolitan statistical areas (MSAs), the analysis computes building damage ratios and flood claim counts by employing k-means clustering for classifying census block groups (CBGs) into distinct property flood risk disposition levels. Then a random forest model is created to specify property flood risk levels of CBGs based on various intertwined hydrological, topographic, and built-environment features. The model transferability analysis results show consistent performance across MSAs, revealing the universality of underlying features that shape city property flood risks. The FloodGenome model is then used to:(1) evaluate the extent to which future urban development would exacerbate flood risk disposition of urban areas; and (2) specify property flood risk levels at finer spatial resolution providing critical insights for flood risk management processes. The FloodGenome model and the findings provide novel tools and insights for improving the characterization and understanding of intertwined features that shape flood risk profiles of cities.