Debris flow susceptibility assessment along the G55 expressway in Guangdong Province, China

Western Guangdong Province in China is frequently affected by intense rainfall and related geological hazards, yet studies on debris flow disasters in this area remain limited. In June 2025, four consecutive debris flow events occurred along the Lianzhou-Sanshui section of the G55 Erenhot-Guangzhou Expressway, attracting significant attention. To better understand the spatial distribution and controlling factors of these hazards, this study conducted a debris flow susceptibility assessment for this section. Based on regional topographic, meteorological, and geological conditions, combined with historical disaster data, an evaluation index system and database were established. Four machine learning algorithms, Back Propagation Neural Network (BPNN), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were applied and systematically compared. Algorithm performance was assessed using 3-fold spatially independent validation. Among the tested algorithms, XGBoost achieved the best predictive performance, with an accuracy of 0.923, recall of 0.879, F1-score of 0.794, and ROC-AUC of 0.927. The debris flow susceptibility was classified into low, medium, and high levels. Medium and high susceptibility areas exhibited clear spatial clustering, mainly in the northern section of the expressway, while the southern section showed generally low susceptibility. Debris flows along the expressway were primarily triggered by shallow surface collapses and landslides that formed cascading hazard chains, while the slipperiness of surface weathering products showed an observed association with debris-flow occurrence and may represent a potential conditioning factor. These findings provide a scientific basis for debris flow risk identification and mitigation and offer insights for improving susceptibility assessment methodologies.