Landslide Hazard Warning Based on Semi-Supervised Random Forest and Effective Rainfall

Accurate early warning of rainfall-induced landslides poses a critical challenge in geological disaster risk management. Conventional deterministic rainfall threshold models often overlook the heterogeneity of regional geological conditions, while landslide susceptibility assessment is plagued by uncertainties in selecting non-landslide samples. To address these issues, this paper took Zhushan County in Hubei Province as the study area, and the semi-supervised random forest (SRF) model was adopted to conduct landslide susceptibility assessment. The critical rainfall (Effective Rainfall-Duration, EE-D) threshold curves were constructed based on the antecedent effective rainfall (EE) and rainfall duration (D). Furthermore, EE-D threshold curves with different geological condition characteristics were established and analyzed according to the thickness, slope, and area of the landslides, respectively. By coupling the landslide susceptibility results with a classified multi-level rainfall threshold model, a spatiotemporally refined regional framework for tiered landslide early warning was developed. The results show that the SRF model solves the problem of non-landslide sample selection error in traditional supervised learning. The Area Under Curve (AUC) value reaches 0.91, which is better than the analytic hierarchy process, logistic regression, etc. Moreover, the models of landslide susceptibility and EE-D threshold can effectively achieve the hierarchical early warning of rainfall-induced landslide hazards.