| In recent years,global extreme weather has occurred frequently,crustal movement has occurred frequently,and the number of geological disasters in China has been high for years,which has caused great harm to the safety of people ’s lives and property.Landslide is the most important geological disaster in China,which has the characteristics of uneven spatial distribution,huge disaster scale and serious damage loss.Landslide susceptibility assessment is a method to analyze and predict the spatial distribution and occurrence probability of landslide disasters.The analysis and prediction results can provide a basis for landslide hazard spatial management,disaster prevention and mitigation policy formulation and national land resource planning.In the study of landslide susceptibility evaluation,sample data is the basis of the model.However,the current selection and optimization methods of sample data are not scientific,resulting in low sample quality.At the same time,the existing susceptibility models retain less spatial information in the evaluation process,resulting in low accuracy of susceptibility evaluation.This study takes Pingwu County,Sichuan Province as the research area,focusing on the scientific issues such as the probability and spatial distribution of landslide disasters in the county.The following studies are conducted :(1)Construct a landslide dataset based on spatial constraints and information quantity constraints.Multi-source geospatial data are collected,11 landslide evaluation factors are selected,and the common line problem among the evaluation factors is analyzed using a geographic probe.The evaluation factors are combined with historical landslide data to analyze the spatial distribution characteristics,genesis mechanism and development environment of landslide points.Based on the existing landslide samples,the density-based spatial clustering of applications with noise(DBSCAN)algorithm and information quantity model are used to determine the spatial constraints and information quantity constraints for the selection of non-landslide samples,and the cosine similarity is used to evaluate the quality of non-landslide samples.The landslide evaluation factor,landslide samples and non-landslide samples together form the landslide dataset.(2)A landslide susceptibility evaluation model based on Vision Transformer(ViT)is constructed.Aiming at the problem that the susceptibility model retains less spatial information in the evaluation process,this study constructs a ViT model based on global feature extraction,and compares the support vector machine model without feature extraction and the residual network model based on local feature extraction.The feasibility of ViT in the field of landslide susceptibility evaluation is verified by the accuracy evaluation index and susceptibility mapping results.It is verified that the selected three models have high accuracy and strong generalization ability.Combined with the field investigation of the study area,the difference of the mapping results is compared,and the global feature extraction method of the ViT model is the best in this study.(3)A composite landslide susceptibility evaluation model combining ViT and Convolutional Neural Network(CNN)is constructed.Aiming at the problem that ViT model cannot accurately capture local features in the evaluation process,this study constructs a composite model of ViT combined with CNN,taking into account the shortcomings of single feature extraction method.The results show that the two composite models have achieved high accuracy and are greater than the single model.By comparing the internal noise of the model,the influence of attention mechanism on CNN and the stability of the model,it shows that the composite model of ViT combined with CNN has the best effect on prediction accuracy,generalization ability and model stability in this study.At the same time,the variation trend of CNN and ViT in the internal noise distribution of the model is compared,which provides a theoretical reference for the future exploration of landslide susceptibility evaluation of ViT combined with CNN composite model. |
