A comprehensive review of key issues in landslide susceptibility prediction and their solutions using semi-supervised imbalanced theory

As the foundation for regional landslide risk assessment, landslide susceptibility prediction (LSP) is a prominent and challenging topic in global landslide disaster prevention and control research. This paper systematically reviews critical issues in LSP, including model selection, identification and integration of conditioning factors, determination and classification of prediction units, methods for selecting non-landslide samples, optimization of the landslide-to-non-landslide sample ratio, assignment of label values to non-landslide samples, and evaluation methodologies for landslide susceptibility results. The literature review indicates that tree-based models, such as Decision Trees and Random Forests, demonstrate superior performance. The selection and integration of conditioning factors should adhere to principles of comprehensive typology and clear physical significance. Prediction units can be defined through multi-scale segmentation of slope units. Non-landslide samples should preferably be randomly selected from areas characterized by very low or low susceptibility. The optimal ratio of landslide to non-landslide samples can be established through experiments under various conditions. Specific small probability values should be assigned to non-landslide samples. Evaluation of LSP results necessitates a comprehensive consideration of multiple metrics, including ROC accuracy, prediction rate accuracy, and the mean and standard deviation of the susceptibility index. Furthermore, to enhance the accuracy of LSP and validate the cross-regional engineering applicability of the semi-supervised imbalanced theory, this study pioneers the application of the Random Forest model based on this theory in Anyuan-Xunwu County, Jiangxi Province. Experiments were conducted under various scenarios, involving different ratios of landslide to non-landslide samples (ranging from 1:1 to 1:260) and varying label assignments for non-landslide samples. Results indicate that as the ratio increases from 1:1 to 1:180, both ROC accuracy and prediction rate accuracy improve gradually from 0.905 and 0.898 to 0.957 and 0.937, respectively. However, no significant improvement is observed in either metric once the ratio exceeds 1:180. Consequently, the optimal ratio of landslide to non-landslide samples is established at 1:180. Additionally, this study collects data from six landslide events that occurred between 2022 and 2024 to validate the results obtained from the 1:180 ratio. The validation reveals that all six landslides are situated in areas of very high or high susceptibility. This not only confirms the effectiveness and engineering applicability of the semi-supervised imbalanced theory in the mountainous and hilly regions of southern Jiangxi but also provides new insights and technical support for the precise prevention and control of landslide risks.