In this study, a machine learning method, i.e., random forest (RF) model, was employed to assess post-fire debris flow (PFDF) susceptibility after the Xichang forest fire occurred on 30 March, 2020. First, we conducted a tracking survey in the rainy season, on-site tests, and remote sensing image interpretation and obtained 10 impacting factors, i.e., the basin area, relief ratio, basin shape coefficient, percentage of area with a slope greater than 50%, proportion of moderate or high severity burned areas, distribution of gravel in the basin, vegetation types and distribution in the basin, early cumulative erosion after fire, peak rainfall in a 1-h interval, and peak rainfall in a 24-h interval, after their correlation test to build a spatial database. Subsequently, a total of 181 PFDF events in the database were randomly divided into training (70%) and validation (30%) samples. Thereafter, the RF model was used to acquire the susceptibility of PFDF in the study area. Finally, receiver operating characteristic (ROC) curve, area under curve (AUC) value, sensitivity, specificity, and accuracy were utilized to validate the predictive performance of the model. Results show that the RF model has good predictive ability with AUC of 93.4%, sensitivity of 88.3%, specificity of 99.3%, and accuracy of 97.8%. This study provides a scientific basis for PFDF disaster prevention and risk management in Xichang City and its surrounding areas.
Read full abstract