Probabilistic flood susceptibility mapping using explainable AI for the Western United States

This study aims to predict flood susceptibility and analyze multi-decadal trends in extreme streamflow events across 1,088 United States Geological Survey flow-monitoring stations. The authors employ an explainable artificial intelligence (XAI) framework, integrating a random forest model with SHapley Additive exPlanations (SHAP), to identify and interpret the dominant drivers of flood susceptibility. The XAI model achieved a high predictive performance (Area Under Curve = 0.92), offering a reliable probabilistic tool for hazard mitigation and emergency planning.

The major findings of the study are as follows:

1. Flood susceptibility hotspots were concentrated along the Pacific slope from coastal Washington through northern California, with a few other regions in interior mountainous regions like the Sierra Nevada and Cascades. Other areas with notable concentrations included the Northern Rockies, Central Rockies, and the Southwest Interior.
2. Low elevation (valley bottoms / coastal plains), high discharge variability, low terrain slope, and statistically significant positive trend in annual mean discharge emerged as the most important predictors of where floods are most likely to occur.
3. Seasonal hydrologic regime that equally reflects streamflow capacity were further important predictors. Low seasonal flow in summer and spring; and high winter flow were associated with heightened flood susceptibility.
4. Finally, for locations identified as flood-prone, we assessed county-level risk by integrating flood susceptibility scores with land cover, human settlement patterns, population exposure and social vulnerability. This produced a spatially explicit composite risk index to guide planning and resource allocation.