Evaluation of highway collapse hazard susceptibility based on the coupling of information value model and multiple machine learning models

Collapse hazards are among the most common geological hazards and pose severe challenges to transportation infrastructure. While machine learning has shown promise in hazard prediction, conventional qualitative methods are often subjective, and single machine learning models are prone to overfitting. To address these limitations, this study developed four coupled models (IV-RF, IV-XGBoost, IV-LGBM, and IV-CatBoost) by integrating an information value model with four machine learning algorithms, using nine influencing factors, such as road distance and precipitation, in Yulin City, China. The susceptibility was classified into five levels: very low, low, middle, high, and very high susceptibility. The AUC values of the coupled models reached 0.963, 0.929, 0.928, and 0.917, respectively, with IV-RF exhibiting the highest predictive accuracy. The proportions of very-high and high susceptibility zones for highway collapses were 31.26% and 44.83%, respectively. These zones were primarily concentrated in western and northwestern Yulin, exhibiting clustering patterns along major highways. The coupled models developed in this study not only provide a direct and reliable scientific basis for collapse risk zoning, key segment monitoring, and targeted engineering mitigation along highways in Yulin but also serve as a valuable reference for assessing highway-related geological hazards in other regions with comparable geographical conditions.