Landslide susceptibility mapping based on stratified non-landslides sampling from engineering geological petrofabric and optimized boosting models: a case of Maiji District, China

Landslide susceptibility mapping (LSM) plays a vital role in preventing and managing geological hazards. The selection of non-landslide sampling methods and model optimization algorithms is essential for LSM.

In this study, stratified sampling was used to construct a non-landslide sample set. Three boosting algorithms, including Categorical Boosting (CB), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM), and three optimization algorithms, including Bayesian Optimization (BO), Random Search Optimization (RSO), and Particle Swarm Optimization (PSO), were combined to construct 12 landslide susceptibility models. The best-performing model was selected, and different classification methods were used to draw the LSM of the Maiji district. In addition, the Shapley Additive exPlanations (SHAP) algorithm was used to reveal the main influencing factors of landslide distribution.

The results showed that the PSO-CB model outperformed other models, with 0.873 in AUC, 80.6% in accuracy, 81.9% in precision, 81.9% in recall, and 81.9% in F1-Score. The LSM revealed that regions with high and very high susceptibility were primarily located in the northwest of the Maiji District, near the Weihe River and its tributaries in the northeast, characterized by frequent human activities. Finally, the PSO-CB model was interpreted using the SHAP algorithm, revealing that land use, elevation, engineering geological petrofabric (EGP), terrain relief, aspect, and rainfall were the primary influencing factors for landslide distribution in Maiji district.

The PSO-CB-SHAP framework can accurately predict landslide-prone zones and explain landslide susceptibility models, providing a scientific basis for the precise prevention and control of landslide disasters in Maiji District.