Quantitative evaluation of uncertainty and interpretability in machine learning-based landslide susceptibility mapping through feature selection and explainable AI

Abstract

Landslide susceptibility mapping (LSM) is essential for determining risk regions and guiding mitigation strategies. Machine learning (ML) techniques have been broadly utilized, but the uncertainty and interpretability of these models have not been well-studied. This study conducted a comparative analysis and uncertainty assessment of five ML algorithms—Random Forest (RF), Light Gradient-Boosting Machine (LGB), Extreme Gradient Boosting (XGB), K-Nearest Neighbor (KNN), and Support Vector Machine (SVM)—for LSM in Inje area, South Korea. We optimized these models using Bayesian optimization, a method that refines model performance through probabilistic model-based tuning of hyperparameters. The performance of these algorithms was evaluated using accuracy, Kappa score, and F1 score, with accuracy in detecting landslide-prone locations ranging from 0.916 to 0.947. Among them, the tree-based models (RF, LGB, XGB) showed competitive performance and outperformed the other models. Prediction uncertainty was quantified using bootstrapping and Monte Carlo simulation methods, with the latter providing a more consistent estimate across models. Further, the interpretability of ML predictions was analyzed through sensitivity analysis and SHAP values. We also expanded our investigation to include both the inclusion and exclusion of predictors, providing insights into each significant variable through a comprehensive sensitivity analysis. This paper provides insights into the predictive uncertainty and interpretability of ML algorithms for LSM, contributing to future research in South Korea and beyond.