Interpretable machine learning incorporating major lithology for regional landslide warning in northern and eastern Guangdong

Landslides pose major risks in northern and eastern Guangdong, China, due to complex geology and heavy rainfall. Traditional models often oversimplify lithology and lack interpretability. This study develops a lithology-specific random forest model that distinguishes between sedimentary and igneous rocks and integrates rainfall, geological, and geotechnical data. Using 754 landslide cases and 1233 rainfall records, the model achieves over 90% hit rate and below 4% false alarm rate. Interpretable machine learning techniques, including feature importance rankings, SHAP values, and partial dependence plots, are used to understand how different factors contribute to landslide occurrence. A case study from the 2024 Pingyuan landslides confirms the model’s real-world applicability. This framework offers improved prediction performance and interpretability and can serve as a robust tool for regional early warning and risk management in geologically diverse areas.