Event-Scale Drivers of Flood Generation: Large-Sample Assessment of the Roles of Soil Moisture and Precipitation

Floods arise from the interaction between near-event precipitation and antecedent soil moisture, which regulates how efficiently precipitation is converted into runoff. Studies often classify floods by generating mechanisms using catchment-level statistics and treat soil moisture implicitly as a bulk proxy, obscuring its direct regulatory role and the influence of specific soil layers on individual flood events. This research utilizes explainable AI tools to quantify the event-scale roles of precipitation and multi-layer soil moisture in flood generation globally. We trained a catchment-shared LightGBM model on the Caravan–GRDC dataset to predict daily streamflow percentiles from 1,385 snow-free catchments. To isolate immediate forcing from antecedent state, we utilized lagged inputs: t−1 for precipitation and t−2 for soil moisture at four depths (0–7, 7–28, 28–100, and 100–289 cm). For 38,317 annual-maximum flood predictions, we applied SHAP to decompose each prediction into predictor contributions and classified events as either precipitation-dominant or soil-moisture-dominant based on the largest absolute SHAP value. Results show that soil moisture is the predominant global flood driver; however, precipitation dominance becomes more frequent toward the highest streamflow percentiles, indicating that the largest peaks often require intense forcing to overcome storage constraints. The two regimes exhibit distinct dynamics: Soil-moisture-dominant floods evolve slowly, with longer rising and recession limbs, and are regulated by shallow subsurface moisture (7–100 cm). They typically occur in larger, flatter, and lower catchments with shallow water tables. Precipitation-dominant floods are flashier, with sharp rising limbs, show stronger sensitivity to the surface moisture (0–7 cm), and are more prevalent in smaller, steeper, high-relief catchments, with deeper water tables. The peak timing predictions reflect the challenge of capturing short-fuse storm dynamics relative to slowly evolving storage states, with well-timed soil-moisture-dominant peaks compared to precipitation-dominant peaks, which exhibited timing delays. This framework provides a scalable, event-based quantification highlighting the catchment control, with soil moisture (especially within the upper meter) acting as an active regulator of runoff generation. The dominance classification supports process-aware forecasting and climate-change adaptation by tracking shifts in flood-generation regimes and indicating whether predictability depends more on storm forcing or antecedent catchment state.