Evaluation of landslide susceptibility of mountain highway based on RF and SVM models.

Geological complexities along mountain highways frequently trigger landslides, posing significant threats to transportation safety and infrastructure. This study evaluates landslide susceptibility along the Lizha-Jiezi section of China's G345 national highway using Random Forest (RF) and Support Vector Machine (SVM) models. Eleven conditioning factors including altitude, slope, aspect, plan curvature, profile curvature, lithology, distance to fault, rainfall, distance to river, normalized difference vegetation index (NDVI), and distance to road were analyzed using remote sensing and field surveys. A landslide inventory of 67 events was divided into training (70%) and validation (30%) datasets, with non-landslide samples selected at least 100 m away from landslide locations to minimize spatial overlap. Factor contribution analysis identified distance to road as the most significant predictor, highlighting anthropogenic impacts on slope destabilization. Model validation via receiver operating characteristic (ROC) curves demonstrated RF's superior performance (AUC = 0.887) over SVM (AUC = 0.735). The RF-derived susceptibility map classified five risk levels, revealing high-risk zones concentrated within 200 m of roads, consistent with field observations. Results emphasize the necessity of integrating anthropogenic factors into landslide risk management for mountainous infrastructure. This study provides actionable insights for mitigation strategies and land-use planning, offering a scalable framework adaptable to similar regions.