Interpretable Coseismic Landslide Susceptibility Assessment Using an Explainable Boosting Machine: A Case Study of the 2022 Luding Earthquake, China

Earthquake-induced landslide susceptibility assessment is important for post-earthquake hazard identification and risk mitigation. However, conventional machine learning models often provide limited insight into their prediction mechanisms. This study applies the Explainable Boosting Machine (EBM) to landslides triggered by the 5 September 2022 Luding earthquake and investigates the nonlinear responses and interactions of the controlling factors. Conditioning factors were screened through correlation and multicollinearity analyses, and the key model parameters were optimized using grid search. The final model included nine main effects and 36 pairwise interactions to characterize nonlinear factor responses and joint effects. The EBM achieved AUC values of 0.987 and 0.976 for the training and testing datasets, respectively, while both Accuracy and F1-score reached 0.920 on the testing dataset. The susceptibility map was classified into five levels using the Natural Breaks method, and the frequency ratio increased from 0.01 in the very low susceptibility class to 6.74 in the very high susceptibility class. Elevation, PGA, and Lithology were the three most important main effects, with importance scores of 1.836, 1.512, and 1.324, respectively. Lithology × PGA and Elevation × Lithology showed the strongest pairwise interactions, indicating that the effects of seismic shaking and elevation on landslide susceptibility were strongly modified by lithological conditions. These results demonstrate that EBM can characterize coseismic landslide susceptibility patterns while directly revealing nonlinear factor responses and pairwise interactions, providing an interpretable framework for understanding the combined influences of seismic, geological, and topographic conditions on landslide susceptibility.