Objectives On July 26, 2024, Typhoon “Gaemi” struck Zixing City, Hunan Province, China, triggering large-scale landslide disasters that caused extensive damage to buildings, blocked several roads, and resulted in significant casualties. However, the development, distribution, and formation conditions of the landslides triggered by Typhoon “Gaemi” remain unclear, which impacts post-disaster reconstruction and risk assessment.

Methods We present the identification and analysis of landslides induced by Typhoon “Gaemi” through three main parts: landslide recognition, spatial and statistical analysis, and explainable machine learning. First, historical rainfall-induced landslide data from regions like Longxi, Yan'an, and Shaoguan are used to train and optimize the YOLOv8 model for improved accuracy and robustness. The model is applied to post-disaster optical remote sensing imagery for automatic identification, with manual verification and the creation of a landslide database for Zixing City. Next, spatial distribution characteristics of the landslides are analyzed using environmental data, revealing the spatial patterns and controlling factors of landslides induced by Typhoon “Gaemi”. Finally, statistical data are used to train a machine learning model to assess landslide susceptibility, with Shapley additive explanations (SHAP) employed to interpret the model and identify key factors influencing landslide occurrence. These methods provide a comprehensive analysis of landslide risks and serve as a basis for disaster prevention.

Results Using the deep learning model YOLOv8 and manual verification, we identify approximately 16 120 landslides triggered by Typhoon “Gaemi” in Zixing City, covering a total area of 32.16 km². These landslides are predominantly concentrated in the eastern mountainous areas of Zixing City, with a peak point density of 129 per km², exhibiting a notable clustering effect. Statistical analysis reveals that most of the landslides occur on slopes with an elevation range of 200⁃600 m and a slope gradient of 20°⁃30°. Further analysis demonstrates that the occurrence of these landslides is significantly influenced by geological conditions, with 63.5% of landslides occurring in granite regions and 29.4% in sandstone and shallow metamorphic rock areas. Random forest modeling and SHAP analysis identify lithology as the primary factor controlling landslide formation. The importance of disaster-triggering factors varies across different lithological regions. In granite areas, faults are the most significant factor influencing landslide occurrence, while human activities also contribute notably to landslide formation. In sandstone regions, river erosion is the dominant factor controlling landslide formation.

Conclusions The deep learning and machine learning models used in this study demonstrate clear advantages in landslide identification, susceptibility assessment, and analysis of disaster-triggering factors. The identified distribution patterns of landslides and the geological control mechanisms can provide scientific evidence for analyzing the landslide formation conditions and disaster prevention and control in Zixing City and similar regions with comparable geological conditions.