Equity-focused flood risk assessment in Maryland using a hybrid explainable machine learning framework with FEMA and USGS data

Abstract Flooding poses a major global threat, causing loss of life and property damage. Accurate flood risk mapping is critical for effective mitigation. Traditional flood zoning approaches often rely on static analyses, failing to capture the dynamic nature of floods primarily driven by changing weather patterns and landscapes. This study develops a hybrid approach integrating a flood inventory from Federal Emergency Management Agency (FEMA) flood zones and U.S. Geological Survey streamflow discharge data with an explainable AI (XAI) model to predict flood susceptibility at the census block group (CBG) scale in Maryland. Trained on a curated flood inventory and environmental variables, the XAI achieves an area under the ROC curve (AUC) and Brier Score of 0.84 and 0.03, respectively. SHAP analysis identifies elevation, distance to water bodies, topographic wetness index, slope, storm frequency, and northern tidewater soils as key contributors to model predictions. Our model shows spatial agreement with FEMA maps (Mantel correlation = 0.72) while capturing further localized hotspots, especially in highland areas. Flood-prone zones (> 70% predicted flood susceptibility) are concentrated along the Chesapeake Bay’s eastern shoreline, including CBGs in Somerset, Dorchester, Queen Anne’s, Talbot, Kent, and Wicomico counties. Caroline County CBGs exhibited the lowest flood exposure. Equity-focused risk assessments incorporating socioeconomic vulnerability and population exposure indicated 422,000 people and $74 million in real estate (based on median home prices) are in high-risk zones. Traditional methods can underestimate these impacts. This hybrid XAI approach provides a scalable framework to enhance probability-based flood risk assessment across neighborhood scales in a changing environment, safeguarding lives and property in flood-prone areas.