Abstract
This study aims to improve landslide susceptibility analysis in Nyingchi City by addressing the challenge of limited landslide sample data. A total of 11 influencing factors-including elevation, slope, aspect, terrain roughness, terrain moisture index, profile curvature, and plane curvature-were initially considered. After correlation and importance analysis, eight key factors were selected for modeling. To augment the limited dataset, the Markov Chain Monte Carlo (MCMC) method was employed to synthetically generate additional landslide sample points. The quality of the generated samples was validated using a Support Vector Machine (SVM) classifier. Further sensitivity analysis and susceptibility modeling were conducted using both the original and augmented datasets. The Light Gradient Boosting Machine (LightGBM) model was selected based on performance evaluation, and its predictive accuracy was assessed using the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. The results show that: (1) The SVM achieved 97.3% classification accuracy with the MCMC-augmented data, indicating the effectiveness of the generation method; (2) The LightGBM model trained on augmented data yielded a higher AUC value than that trained on original data; (3) The most influential factors for landslide susceptibility were distance to roads, aspect, and elevation; (4) Although there were minor differences in the susceptibility maps generated from the two datasets, their overall spatial patterns were similar. This study provides methodological insights for landslide risk assessment and disaster mitigation in resource-limited mountainous areas.