Objective The Panzhihua-Xichang (Panxi) region generally slopes from the northwest to the southeast. Characterized by intricate geological structures and intense seismic activity, the area also experiences concentrated rainfall with frequent rainstorms. These conditions lead to frequent debris flow disasters, which pose severe threats to the lives and property of local residents. Therefore, it is imperative to conduct scientific debris flow hazard assessments. In such assessments, the classification intervals of continuous evaluation factors and the selection of machine learning models are critical determinants of result accuracy. These findings ultimately provide a scientific basis for debris flow forecasting and prediction.

Methods In this study, sub-watersheds were adopted as evaluation units. The geodetector method was employed to establish the relationships between evaluation indicators and actual debris flow points, exploring the similarity between different classification intervals of continuous evaluation factors and the spatial distribution of debris flow disasters. Based on this, the classification intervals of the evaluation factors were optimized, and the evaluation accuracies of the random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) models were compared.

Results 1) The RF model using classification intervals optimized by the geodetector method achieved the highest accuracy, with an AUC of 0.935. This performance was superior to that obtained using the conventional natural breaks method, indicating that geodetector-based classification optimization significantly improved the accuracy of debris flow hazard assessment. 2) All three machine learning models exhibited good predictive performance and showed high consistency in identifying the main disaster-causing factors of debris flows in the Panxi region. Elevation was identified as the most important factor, followed by the distance to water system, annual average rainfall, and maximum annual rainfall. This indicates that topographic and rainfall factors are critical triggers for debris flow disasters. Owing to its strong nonlinear modeling capability, the RF model outperformed the XGBoost and SVM models in handling complex multi-factor interactions. Specifically, the RF model achieved an AUC of 0.935 and an accuracy of 0.854, which were higher than an AUC of 0.926 and an accuracy of 0.842 for the XGBoost model and an AUC of 0.921 and an accuracy of 0.829 for the SVM model, demonstrating its overall superior performance.

Conclusions Based on the geodetector, the classification intervals of continuous evaluation factors can be effectively optimized, thereby improving the accuracy of the assessment results. The RF model achieved the highest accuracy in the debris flow hazard assessment of the Panxi region. The evaluation results can provide a more scientific theoretical basis for debris flow disaster prevention and mitigation in the Panxi region.