Interpretable and Calibrated XGBoost Framework for Risk-Informed Probabilistic Prediction of Slope Stability

This study develops an interpretable and calibrated XGBoost framework for probabilistic slope stability assessment using a 627-case database of circular-mode failures. Six predictors, namely, unit weight (γ), cohesion (c), friction angle (φ), slope angle (β), slope height (H), and pore-pressure ratio (rᵤ), were used to train a gradient-boosted tree model optimized through Bayesian hyperparameter search with five-fold stratified cross-validation. Physically based monotone constraints ensured that failure probability (P_f) decreases as c and φ increase and increases with β, H, and rᵤ. The final model achieved strong performance (AUC = 0.88, Accuracy = 0.80, MCC = 0.61) and reliable calibration, confirmed by a Brier score of 0.14 and ECE/MCE of 0.10/0.19. A 1000-iteration bootstrap quantified both epistemic and aleatoric uncertainties, providing 95% confidence bands for P_f-feature curves. SHAP analysis validated physically consistent influence rankings (φ > H ≈ c > β > γ > rᵤ). Predicted probabilities were classified into Low (P_f < 0.01), Medium (0.01 ≤ P_f ≤ 0.10), and High (P_f > 0.10) risk levels according to geotechnical reliability practices. The proposed framework integrates calibration, uncertainty quantification, and interpretability into a comprehensive, auditable workflow, supporting transparent and risk-informed slope management for infrastructure, mining, and renewable energy projects.