Abstract
Landslides are a prevalent and devastating form of geological disaster. These events occur when gravity causes rock and soil masses to slide along specific surfaces or zones, often triggered by intense rainfall, seismic activity, or human engineering activities. Assessing landslide hazard risk is crucial for effective disaster management, yet traditional approaches often rely on administrative or grid units, which lack the granularity needed for site-specific hazard management. This results in uniformly high-risk classifications for hilly areas, complicating practical engagement and increasing management costs. The study further combines historical landslide data and applies machine learning models such as Random Forest, XGBoost, and LightGBM to analyze landslide susceptibility in Wenzhou City, proposing a slope unit-based landslide hazard assessment method. The results are as follows: (1) Landslide Susceptibility across different slope units was categorized as low, low-moderate, moderate, moderate-high, high, and very high, with the very high-risk slope units accounting for 5.35% of the total area and the low-risk slope units covering the largest area (975.41 km(2)). (2) Among the machine learning models used for landslide susceptibility analysis at the slope unit level, the Random Forest model performed the best, demonstrating higher prediction reliability, with an accuracy of 77.94% for Random Forest, 76.95% for XGBoost, and 78.30% for LightGBM. (3) Extreme rainfall events significantly increased the proportion of high-risk slope units, particularly in mountainous and hilly areas. According to different rainfall return periods, the proportion of very high-risk slope units increased from 5.35 to 40.39% under the 100-year return period. (4) A case study of Xuekou Village validated the practical application of the slope unit risk assessment results and proposed preventive measures for medium-to-high-risk units, such as regular monitoring and enhanced vegetation coverage, to mitigate landslide risks.