From Rainfall to Absolute Landslide Probability: A Machine Learning–Based Approach

Abstract Landslides are prevalent geological hazards in mountainous and hilly regions, posing significant threats to human lives, property, and infrastructure, especially when external factors such as heavy rainfall elevate the risk of occurrence. In recent years, the rapid development of remote sensing technology, geographic information system, and big data analytics has enabled the establishment of landslide databases and the refinement of landslide hazard assessment methodologies, opening new avenues for disaster prediction and prevention research. This study focuses on the probabilistic hazard assessment of rainfall-induced landslides, using the heavy rainfall event in Nanping City, Fujian, in 2019 as a case study. The absolute probability of landslide occurrence was quantified by integrating machine learning models with Bayesian probability theory. To analyze their contributions to landslide hazard, the study comprehensively considered nine key factors: elevation, slope angle, slope aspect, topographic wetness index, fault distance, rainfall amount, topographic position index, land cover, and stratum. The findings reveal that rainfall intensity is the primary trigger for landslides. However, the occurrence of landslides also depends on the combined influence of topography, geology, and human activities rather than rainfall alone. Building on these insights, this study extends the established landslide probability assessment model to a wider region. And simulating landslide occurrence probabilities under various rainfall scenarios. This approach confirms the strong correlation between rainfall and landslides while offering scientific support for emergency management and disaster warning systems. The research provides a robust foundation for disaster prevention and control in mountainous areas, infrastructure planning, and the development of emergency response systems, particularly by refining landslide prediction models that account for multiple factors. Additionally, the findings underscore the need to dynamically adjust and optimize these models for future landslide predictions, incorporating regional characteristics to enhance the accuracy and timeliness of early warnings.