Application and interpretability of ensemble learning for landslide susceptibility mapping along the Three Gorges Reservoir area, China

Landslides pose a significant threat to China’s Three Gorges Reservoir area. Many ensemble learning models have been applied to landslide susceptibility mapping (LSM) in this region, as it forms the foundation of landslide risk management. However, most landslide susceptibility models lack interpretability, hindering the explanation of the relative importance and interactive mechanisms among landslide conditioning factors. This study evaluates and interprets three tree-based ensemble learning models—XGBoost, Random Forest (RF), and Light GBM—for LSM in the Yichang section of the Three Gorges Reservoir area, employing SHAP (SHapley Additive exPlanations) analysis. Among these models, XGBoost and RF exhibit similar the area under the receiver operating characteristic curve (AUROC) values of 0.96 and 0.95, slightly outperforming Light GBM with an AUROC of 0.93. We identify four crucial landslide conditioning factors from a dataset of 714 landslide samples by individual interpretation, shedding light on specific elements that drive higher susceptibility and recommending suitable mitigation measures for different landslide. Global interpretation via SHAP reveals that elevation, Normalized Difference Vegetation Index, distance from river, distance from road, slope, and lithology are the primary factors influencing landslide susceptibility. We delve deeply into the relationships among these factors, their values, and the mechanisms triggering landslides. In addition, to enhance the credibility and reliability of SHAP interpretation results, we cross-referenced these results with relevant literature on the formation mechanism of landslides in the Three Gorges Reservoir area. This study contributes to a better understanding of landslide risk management and bridges the gap between advanced machine learning models and interpretable results by introducing SHAP. Furthermore, we augment the SHAP analysis results with domain-specific expertise in the field of landslides, helping to bridge the potential shortcomings of SHAP as a data-driven-based approach.