Enhancing Landslide Susceptibility Assessment With a Novel IV ‐ LR ‐ RF Coupled Model: A Comparative Study in Lvhe Town, Qinling‐Daba Mountains, China

The Qinling‐Daba Mountains, characterized by active tectonics, rugged topography, and intense human activities, are among the most landslide‐prone regions in China. Accurate landslide susceptibility assessment (LSA) is critical for regional hazard mitigation; however, standalone machine learning models often suffer from insufficient accuracy and robustness in township‐scale studies with limited historical landslide samples. This study aims to develop a novel multi‐stage coupled model to overcome this limitation and provide refined LSA results for local disaster management. Taking Lvhe Town in Xunyang City, Shaanxi Province, as a case study, we conducted LSA at the slope‐unit scale using the proposed IV‐LR‐RF three‐stage coupled model. This model sequentially integrates the information value (IV) method for objective factor quantification, logistic regression (LR) for physically interpretable linear weighting, and random forest (RF) for robust nonlinear classification. The performance of the IV‐LR‐RF model was systematically compared with three benchmark models (standalone RF, IV‐RF, and LR‐RF) through a validation framework that included confusion matrix metrics, ROC curves, and spatial consistency analysis with landslide inventories. The results show that all coupled models outperform the standalone RF model, with IV‐LR‐RF achieving the best overall performance—a test set accuracy of 0.967 and an AUC (area under the curve) of 0.937. Notably, the high and moderate susceptibility zones, which account for only 30.39% of the study area, contain 77.1% of historical landslides and all four new landslides triggered by extreme rainfall in 2025. The zoning results also exhibit high spatial consistency with the official 2024 geological hazard risk survey of Shaanxi Province. This study demonstrates that the multi‐stage coupling strategy effectively combines the strengths of traditional statistical methods and machine learning, significantly improving LSA accuracy under limited sample conditions. The proposed model provides a robust scientific basis for precise landslide prevention, territorial spatial planning, and emergency investigation deployment in complex mountainous regions.