Explainable tree-based machine learning models with Bayesian optimization in Optuna for flood susceptibility assessment

Floods are among the most destructive natural hazards, particularly in regions with complex terrain and hydrology, where they cause substantial environmental, social, and economic losses. Reliable flood susceptibility mapping is therefore essential for disaster risk reduction and mitigation strategies. Here, we assess flood susceptibility in Charsadda, Pakistan, by integrating machine learning with explainable artificial intelligence (XAI). Seven tree-based ensemble algorithms were trained on twelve conditioning factors and a historical flood inventory. Hyperparameters were optimized using Bayesian optimization in the Optuna framework, and SHAP (SHapley Additive exPlanations) was applied for interpretability. All models achieved high predictive performance (AUC > 0.95), with Gradient Boosting (GB) producing the best overall classification accuracy (accuracy, precision, recall, and F1-score ≈ 0.93; AUC = 0.962), whereas ExtraTrees attained the highest AUC (0.977). The GB identified 19.3% and 13.0% of the study area as very highly and highly flood-prone, mainly along the Kabul and Swat river corridors. SHAP analysis identified elevation, height above nearest drainage (HAND), drainage density, NDVI, lithology, and land use/land cover as the dominant controls on flood occurrence. By combining advanced optimization and model interpretability, this framework enhances predictive accuracy and transparency, providing more reliable and actionable insights for disaster preparedness and risk mitigation in data-scarce regions.