Statistical knowledge-guided multi-classification machine learning modeling scheme for landslide susceptibility levels

This study proposes a novel multi-classification Machine Learning (ML) scheme for Landslide Susceptibility Modeling (LSM) to address limitations in binary classification models. Traditional LSM often uses binary samples (landslide vs. non-landslide) and applies classification models for regression prediction, leading to information loss and spatially discontinuous mapping results. The proposed scheme integrates statistical knowledge-guided multi-classification modeling with four key steps: (1) generating an initial Landslide Susceptibility Zonation Map (LSZM) using mathematical statistics as prior knowledge; (2) refining the landslide inventory to establish very high susceptibility level samples; (3) determining sample sizes for other susceptibility levels based on zonal area ratios; and (4) sampling within corresponding zones. Experimental results in Lvliang City, Shanxi Province, using 8 ML algorithms, demonstrate that multi-classification modeling outperforms binary modeling and the initial LSZM in terms of the Hit Optimization Index (HOI), with an average improvement of 6.71%. HOI quantifies historical landslide coverage within minimized very high susceptibility zones, reflecting real-world accuracy. SHAP-based explanatory analysis identified elevation, road density, and average annual rainfall as key influencing features, aligning with local geological survey findings. This scheme enhances the accuracy and spatial continuity of LSZM, providing critical support for risk assessment and mitigation. The multi-classification scheme effectively reduces prediction errors and improves the representativeness of susceptibility levels, offering a robust framework for LSM. • A multi-class modeling scheme addressing info loss in binary models is proposed. • A landslide hit optimization index for assessment of mapping results is designed. • The multi-class modeling outperforms binary modeling in the mapping results. • Explanatory analysis results are highly compatible with the local geologic surveys.