Analysis Of Debris Flow Susceptibility In Shigatse Area Based On Machine Learning Algorithm

China's mountainous area is close to 70%,and geological disasters occur frequently.Debris flow is a common geological disaster,which is the gravity-driven movement of a solidliquid mixture,causing huge economic losses and casualties.Debris flow susceptibility analysis has important guiding significance for early warning and management of mountain disasters.Debris flow susceptibility analysis can predict the probability of debris flow occurring based on regional environmental factors.In this paper,the Tibet Shigatse region with frequent debris flow is selected as the research area.The satelite remote sensing data and debris flow event history data are combined with the Geographic Information System(GIS)and machine learning algorithms to study the debris flow susceptibility,which provides a scientific basis for urban construction and planning in Shigatse.Firstly,the 16 environmental factors affecting the occurrence of debris flow are obtained and calculated,and the different attributes of multi-source data are unified based on GIS.For the first time,the Normalized Difference Laten Heat Index(NDLI)is applied to the debris flow susceptibility analysis.The relationship between environmental factors and debris flow distribution was analyzed according to the frequency ratio method,and the results were used as the input of the Self Organizing Maps(SOM)neural network to analyze the debris flow susceptibility,and generate "non-debris flow" unit.The verification results show that it is feasible and reasonable to analyze the susceptibility of debris flow based on clustering method.Then,the "non-debris flow" unit is selected from the SOM-based debris flow susceptibility map and is used together with the debris flow unit as input to five machine learning methods(BPNN,one-dimensional CNN,DT,RF and Extreme Gradient Boosting(XGBoost).This paper combines SOM and XGBoost for the first time and applies it to the study of debris flow susceptibility.The results of five model evaluation methods(Precision,Accuracy,Recal,F1 score and AUC)show that The SOM-XGBoost model has the best test accuracy: XGBoost(0.953)> RF(0.943)>1D-CNN(0.939)>BPNN(0.932)>DT(0.898),and the prediction results of the study area have a good correlation with the debris flow distribution,and the prediction efficiency is high: DT(0.7)>XGBoost(4.5)>BPNN(7.6)>1D-CNN(8)>RF(10.8)(unit: min).Finally,sorting 16 environmental factors based on the the tree's “feature importance”.The experimental results show that NDLI ranks first,and the other three main factors are annual average rainfall,profile curvature,and plane curvature.