Enhancing Landslide Susceptibility and Dynamic Exposure Assessment Using Interpretable Machine Learning: A Case Study of the Qinba Mountain Area, China

ABSTRACT Landslide susceptibility assessment is a critical step in preventing and mitigating landslide disaster risks, providing a basis for avoiding potential landslide hazards. Although machine learning has been widely adopted in this field, existing studies often rely on single models without rigorous comparative validation and frequently overlook the spatiotemporal dynamics of exposed elements at risk, limiting the accuracy and practical applicability of risk assessments. Addressing these gaps, this study employs the Qinba Mountain Area, a region prone to landslides, as a case study to integrate multi‐model susceptibility mapping with dynamic exposure analysis. Four machine learning methods, Random Forest (RF), Multivariate Adaptive Regression Splines (MARS), Generalized Additive Model (GAM), and Support Vector Machine (SVM), are used to construct landslide susceptibility models for the region, using 18 conditioning factors. The results indicate that the RF model achieved the highest predictive accuracy (AUC = 0.815), outperforming MARS (0.765), GAM (0.763), and SVM (0.760). The susceptibility map derived from the optimal RF model reveals that high‐risk zones exhibit a clustered distribution in the mountainous terrains of Nanyang, Hanzhong, and Ankang. Furthermore, by integrating time‐series GDP and population grid data (2000–2010), this study uncovers a significant expansion of exposure in high‐susceptibility zones due to urbanization. These findings demonstrate the necessity of coupling susceptibility modeling with dynamic exposure analysis, providing a scientific basis for spatial planning, early warning systems, and sustainable urban development in complex mountainous regions.