Susceptibility Assessment of debris flow disaster based on machine learning models in the loess area of Yili valley

The Yili Valley, located on the border between China and Kazakhstan, serves as the juncture of North and South Xinjiang, and stands as a pivotal outpost on the Silk Road Economic Belt. This area possesses a fragile ecological environment and experiences frequent debris flow disasters. In this study, four machine learning models--Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Decision Tree (DT)-- were employed to evaluate the debris flow susceptibility and compute the weights of evaluation factors. The models were fed inputs comprising 398 identified debris flow channels and 14 feature parameters such as fault density, topographic relief, land use, ndvi, multi-year average rainfall, etc obtained through remote sensing and field surveys. Also, the accuracy of the four machine learning models was evaluated by ROC curves and calculating the Area Under the Curve (AUC). The research results show that: 1. High debris flow susceptibility areas are mainly located in the Tianshan Mountains in the deep river valley region and the loess-covered areas in the mountain front slopes; 2. Multi-year average rainfall, drought index, and topographic relief variability are the top three influential factors controlling the spatial development of debris flows; 3. The AUC values for the validation datasets of the four models were 0.79 (DT), 0.89 (LR), 0.938 (RF) , and 0.932 (SVM), with the Random Forest model exhibiting superior predictive capability in assessing susceptibility in the region; 4. The disruption of ecological vegetation in the loess-covered region of the study area is a significant cause of frequent debris flow occurrences. Ecological governance and protection efforts should be emplasized to reduce soil erosion and effectively mitigate debris flow disasters at their source.