Accurate landslide susceptibility mapping (LSM) is critical for disaster risk assessment and mitigation, particularly in regions vulnerable to rainfall-induced slope failures. While sample selection is a key determinant of LSM accuracy, most existing approaches rely on landslide centroids or entire polygons due to resolution constraints. The potential advantages of boundary-based sampling within source areas remain largely unexplored. This study investigates the rainfall-triggered landslides that occurred in July 2013 in Tianshui City, China. Using high-resolution remote sensing and LiDAR data, we established a detailed inventory of landslides and their source areas. We developed a multi-dimensional LSM framework and introduced a novel Landslide Susceptibility Score (LSS), derived from the confusion matrix, to quantitatively assess the accuracy and practical reliability of susceptibility maps. The findings demonstrate that boundary sampling within source areas substantially enhances predictive performance. The Random Forest model achieved the best results (AUC = 0.960, F1 = 0.890, LSS = 37776), validating the methodological advantage of this sampling strategy. Furthermore, SHapley Additive exPlanations (SHAP) were employed to construct a multidimensional reasoning framework for landslide failure mechanisms, revealing the dominant roles of slope aspect, elevation, and surface roughness in modulating hydrological processes that drive slope instability. This research highlights the geotechnical and geomorphological significance of boundary-based source area sampling, establishing a robust methodological foundation for high-precision and automated LSM under extreme rainfall conditions. The proposed LSS further provides a comprehensive and scalable tool for evaluating susceptibility model performance in future studies.
扫码关注我们
求助内容:
应助结果提醒方式:
