Landslide susceptibility modeling based on SHAP interpretability and ensemble learning: a case study in Fuyuan County, Southwest China

Landslides cause severe ecological, human, and economic losses globally, with Fuyuan County in Yunnan Province, China, being a typical case. Accurate landslide susceptibility mapping (LSM) is crucial for disaster prevention and mitigation. Traditional methods struggle to meet contemporary needs, so this study employs advanced integrated machine learning models (LightGBM and XGBoost) to assess landslide susceptibility in the county, comparing them with traditional models. The LightGBM model performed best, achieving an AUC of 0.89, accuracy of 84.10%, and F1-score of 83.17%. It also demonstrated excellent stability—characterized by low uncertainty, narrow confidence intervals, and consistent discriminative ability across multiple resamplings—with outstanding reliability. Frequency ratio analysis identified key factors facilitating landslides: weak mudstone lithology, proximity to rivers (<200 m), and high mining density (0.098–0.149). Dense vegetation and hard limestone, however, reduce landslide risks. SHAP analysis further revealed that mining density is the most significant influencing factor, with a synergistic effect with river proximity that jointly exacerbates landslide susceptibility. The generated susceptibility zoning map identifies most areas of Dahe Town, Yingshang Town, and Zhuyuan Town as very high susceptibility zones, which highly aligns with historical landslide records and field survey results. The study emphasizes the need to strictly restrict mining activities in very high susceptibility zones and riparian areas, while comprehensively implementing slope reinforcement, vegetation restoration, and mine reclamation measures. These findings provide a scientific theoretical framework for global landslide research.