Flash floods constitute an escalating hazard in the Texas Hill Country, where steep topography and hydroclimatic extremes converge to produce rapid-onset disasters. In July 2025, Kerr County experienced a catastrophic event following more than 380 mm of rainfall within 24 hours, underscoring the urgent need for predictive frameworks that are both accurate and physically interpretable. Here, we develop an explainable machine learning (XAI) framework to model flash-flood susceptibility in the Guadalupe River Basin, integrating high-resolution geospatial predictors including Euclidean distance to streams, valley depth, digital elevation model (DEM), LS-factor, CHIRPS-based seven-day precipitation, topographic wetness index (TWI), and curve number (CN). A balanced dataset of 2,000 pixels was used to train eight classification models; CatBoost delivered the best performance (AUC = 0.93, F1 = 0.89), surpassing XGBoost, Random Forest, and neural networks. By combining SHAP (SHapley Additive exPlanations) and counterfactual analysis, we revealed threshold-dominated flood drivers. Terrain-related variables—particularly stream proximity and valley depth—exerted nonlinear, dominant control on flood susceptibility, with critical transitions emerging below 150 m of stream distance and above 20 m of valley depth. Counterfactuals confirmed that small perturbations in these variables alone were sufficient to reverse predicted flood outcomes, delineating the decision boundaries of the model and enabling physically grounded diagnostics. This integrative XAI framework offers a transferable, physics-aware methodology for identifying dominant controls on flash-flood risk, advancing both scientific understanding and operational readiness in data-scarce, topographically complex environments.
Alcântara et al. (Sun,) studied this question.