Landslide susceptibility analysis for Ancient Shu Roads using ensemble machine learning

Landslides pose continuous threats to linear cultural heritage systems constrained by complex topography and geomorphic processes. Taking China’s Ancient Shu Roads as a case study, this study establishes a landslide susceptibility assessment framework for linear cultural heritage. Slope units and pixel units are adopted, with 15 factors covering terrain, geology, vegetation, hydrology, and human activities. Ensemble machine learning models are used to reveal nonlinear relationships between landslides and influencing factors. Results show that the TabPFN model yields the optimal performance with an accuracy of 80.38%. The Tangluo Route has the lowest susceptibility, while the Baoxie and Chencang Routes show higher values. Finer-scale slope units (SUs2) provide the best spatial representation. Distance to rivers (DRI), slope angle (SA), and rainfall (RA) are the dominant driving factors. High landslide susceptibility mainly occurs in rainy seasons, especially from March to September.