As landslides are typically rare events, conventional negative sample selection methods are prone to incorporating latent hazard pixels into the non-landslide sample pool during blind random sampling across a region, thereby leading to negative sample contamination. Therefore, focusing on Zanda County, Xizang, this study first constructed a high-precision hazard database based on remote sensing interpretation and field verification. A benchmark random forest (RF) model was then utilized to isolate high-risk hazard zones, allowing non-landslide points to be randomly sampled within explicitly secure areas. This process established an adaptive negative sample optimization strategy for landslide susceptibility modeling that balances sample reliability with global statistical distribution rationality. Subsequently, landslide susceptibility was evaluated using multiple advanced machine learning architectures, including CatBoost, TabNet, and LightGBM. Finally, a multi-dimensional verification framework combining the statistical distribution characteristics (mean and variance) of the Landslide Susceptibility Index (LSI) with Area Under the Curve (AUC) values was developed to validate the spatial rationality of the output maps and facilitate their engineering application. Results demonstrate that the proposed modeling approach aligns rigorously with the actual spatial rare-event distribution logic of landslides, effectively balancing local prediction accuracy with global spatial plausibility. Among the evaluated models, the coupled RF-CatBoost model exhibited significant superiority in reducing false positive rates within safe zones and eliminating abnormal probability fluctuations along the highway corridor. The resulting susceptibility maps provide highly practical intelligent technical support and a scientific basis for the refined protection of linear infrastructure in Zanda County and similar alpine canyon regions.
 

Landslide prediction；Coupled models；Sample optimization；Highway corridor；Qinghai-Xizang Plateau