Interpretable Flood Probability Modelling from Multidimensional Drivers Using Two-Stage Feature Selection and Hybrid Optimised Machine Learning

Abstract Floods are among the most destructive natural disasters, driven by environmental, infrastructural and socio-political factors. This paper presents an interpretable framework for estimating flood probability from 20 input variables in a benchmark dataset of 50,000 records, covering hydroclimatic and physical, land-use and ecosystem, infrastructure and water-management, and socioeconomic, governance and planning factors. To overcome the black-box nature of machine-learning models, the framework combines two-stage feature selection, in which Variance Inflation Factor (VIF) screening checks for collinearity and Recursive Feature Elimination (RFE) removes redundant predictors, with SHapley Additive exPlanations (SHAP)-based sensitivity analysis. Decision Tree Regression (DTR), Support Vector Regression (SVR) and Extreme Gradient Boosting (XGBoost) serve as base regressors, with Lungs Performance-Based Optimisation (LPO) used for hyperparameter tuning. VIF screening showed no severe multicollinearity (VIF < 6), so all 20 variables were retained, and RFE then retained the 18 highest-ranked predictors for each regressor. XGBoost was the best base regressor. Coupling each regressor with LPO improved test-set performance, increasing R² by 11–17% and reducing the root mean square error (RMSE) by 34–61%. The LPO-optimised XGBoost hybrid (XGLP) performed best (test R² = 0.97; RMSE = 0.0089), a 17% increase in R² and a 61% reduction in RMSE relative to the base XGBoost. SHAP analysis showed near-uniform contributions across predictors, consistent with the equal weighting of drivers in the benchmark index. The framework offers a transparent, reproducible pipeline applicable to observed flood records to support climate-resilience and infrastructure planning.