Quantifying the Relative Contributions of Different Flood Generating Mechanisms to Floods across CONUS

Floods in a single catchment may be attributed to multiple flood generating mechanisms. Few studies made quantitative investigation about the relative contributions of multiple mechanisms to local floods. Here we used interpretable machine learning to investigate the extent to which different flood generating mechanisms contribute to local floods across the Continental United States (CONUS). Four flood generating mechanisms were first formulated in this study, including (1) precipitation (P), (2) precipitation excess (Pe), considering precipitation infiltration upon unsaturated soil, (3) snowmelt (Ps), (4) effective precipitation (Peff), considering both snowmelt and infiltration. Tree-based regression models (XGBoost, random forest) combined with interpretable machine learning (SHAP, ALE) were used to quantify the relative contributions of these mechanisms to floods. The relative contributions from these mechanisms display clear regional patterns. In the west coast, Pe and Peff (~0.4) have much higher contributions to floods than P and Ps (~0.1), indicating infiltration is nonnegligble process in both rainfall- and snowmelt-caused floods. In the Rocky Mountains, Ps contributes the highest to local floods (~0.5), followed by P (~0.3). Pe largely shapes the floods in the southeastern CONUS (~0.4). Peff is dominantly the influential mechanism in the midwest (~0.4), indicating both snowmelt and infiltration processes should be accounted for in this region. The conclusions are independent of machine learning model choices, and are in line with the qualitative assessment in terms of co-occurrence probability and magnitude correlation between precipitation variants and floods, indicating the robustness of the quantitative results.