Coupling hyperparameter optimization and explainable artificial intelligence (XAI) for high-precision landslide susceptibility mapping in complex mountainous environments

Landslide susceptibility mapping (LSM) is critical for disaster risk mitigation in seismically active and topographically complex regions. Existing studies inadequately address class imbalance and model interpretability, particularly in the Nujiang Lisu Autonomous Prefecture—a region of deeply incised valleys and active tectonics. This study proposes a hybrid framework integrating Particle Swarm Optimization (PSO) with Random Forest (RF) and a balanced negative sampling strategy (1:1 ratio, pairing documented landslides with 9,700 non-landslide cells outside a 500 m buffer). Using a multisource spatial database, PSO-tuned RF outperformed SVM and XGBoost benchmarks, achieving the highest accuracy (0.860) and AUC (0.928). SHAP analysis identified elevation, soil type, and rainfall as dominant predictors. The resulting high-resolution susceptibility map reveals high-risk zones along Nujiang River flanks and areas of intense anthropogenic activity. By coupling hyperparameter optimization with model interpretability, the framework offers regionally grounded support for hazard prevention and land-use planning in the Nujiang Lisu Autonomous Prefecture.