Evaluation Of Town-level Landslide Hazards Susceptibility Based On Multiple Evaluation Units And Evaluation Models

Xuanwo Township is located in the southwest of Hanyin County,Ankang City,Shaanxi Province,with complex geological and environmental conditions and frequent human engineering activities in the area,which is one of the high incidence areas for geohazard breeding.In this thesis,based on data collection and field survey,Arc GIS,SPSS,Python and other software are used as research tools,raster cells and slope cells are selected as evaluation units,and a deep learning model(CNN model)and machine learning models(RF model,ANN model and XGBoost model)are established for landslide hazard susceptibility evaluation in the study area,and the main results are as follows :(1)There are 127 landslide hazards in the study area,mainly small mound landslides;the plane form is mainly tongue-shaped and the profile form is mainly linear;landslides in the area are mainly concentrated in July-August,and are mostly developed in the gently sloping areas of low hills in the north-central part of the country where the mountains are low,the water system is developed and the population is dense;(2)Two kinds of 12.5×12.5m grid elements and slope elements were selected as evaluation units;The selected evaluation factors were analyzed and screened by Pearson correlation coefficient method and information gain method,and finally the relative height difference,aspect,slope,curvature,TWI,NDVI,stratigraphic lithology,rock and soil type,distance from structure and distance from road were retained A total of 12 evaluation factors were used to construct the evaluation index system of landslide disaster susceptibility in the study area.(3)Using Arc GIS software and Python language,two deep learning models(CNN model)and traditional machine learning models(RF model,ANN model,XGBoost model)for evaluation of landslide susceptibility in the study area were established,and the evaluation results were divided into very high,high,medium,low,and very low,susceptibility zones based on the natural interruption point method 5 parts.(4)The rationality test was used to verify and analyze the evaluation results of different evaluation models of various evaluation units based on the ROC curve accuracy test and the experience of three-dimensional view accuracy.The results show that the deep learning algorithm is better than the traditional machine learning model,and the success rate and prediction ability of the CNN model are the highest,and the AUC value of the training set of the CNN model is when the raster cell and the slope element are used for susceptibility evaluation 0.99,the prediction accuracy AUC values were 0.90 and 0.91,the zoning results were the most reasonable.Among them,the area share of the CNN model based on raster cells for the partitioning of susceptibility is 6.53% for the very high susceptibility area,16.96% for the high susceptibility area,28.87% for the medium susceptibility area,28.22% for the low susceptibility area,and 19.41% for the raster very low susceptibility area;the area share of susceptibility zoning based on slope cells was 11.39% for very high susceptibility,19.09% for high susceptibility,23.99% for medium susceptibility,25.31 for low susceptibility,and 20.22 for very low susceptibility.