Abstract
Flash floods endanger communities and ecosystems in rugged regions, but precise prediction is difficult due to environmental complexity. This study evaluates six machine learning algorithms for flash flood mapping in Iran’s Dez Basin, a region growing more vulnerable to climate extremes. We developed an integrated geospatial database incorporating 32 climatic, anthropogenic, and physiographic parameters, validated through extensive field surveys documenting historical flood events. The dataset (70% training, 30% validation) was analyzed using: (1) H2O Deep Learning framework, (2) Random Forest (RF), and (3) four boosting methods (AdaBoost, XGBoost, LightGBM, CatBoost). The RF model achieved exceptional predictive performance (AUC = 0.89, accuracy = 95%), outperforming other techniques by 6–12% in classification metrics. Sensitivity analysis identified precipitation intensity (β = 0.34, p < 0.01), watershed area (β = 0.28), and slope gradient (β = 0.25) as statistically significant dominant controls.These findings advance flood risk management in three key ways: First, they demonstrate RF’s superiority in handling heterogeneous geospatial data. Second, the 30 m-resolution susceptibility map provides actionable insights for land-use planning. Third, the methodology offers a transferable framework for arid/semi-arid regions globally. We recommend policymakers prioritize slope stabilization and early-warning systems in high-risk zones (AUC > 0.85) to enhance community resilience.