Spatial Prediction And Risk Assessment Of Landslides Based On Machine Learning

In recent years,there are many geological disasters in China.This has a serious impact on social economy and urban development,and landslides is the most frequent and harmful geological disasters.Therefore,people are more and more aware of the importance of preventing and reducing landslides for economic development and sustainable development.Therefore,by studying the development characteristics of landslide disaster and scientifically assessing the risk of landslide disaster,we can quickly and accurately predict the potential loss of disaster,delimit the risk area,and provide important decision support and scientific basis for disaster prevention and mitigation.In this thesis,Tongjiang County of Bazhong City is taken as the research area,and based on machine learning and GIS technology,the landslide evaluation factors processing,landslide prone spatial prediction,Pu bagging model,cellular automata Markov model(abbreviated as cellular automata Markov model)are carried out Starting with the research of Ca Markov),different methods of unit division and different basic learning devices are used to build the spatial prediction model of landslide susceptibility.The spatial prediction of landslide susceptibility based on Pu bagging model and the dynamic prediction of landslide susceptibility based on Ca Markov model are studied.A dynamic prediction model based on Ca Markov with high accuracy is established through several groups of comparative experiments.Combined with the density analysis of the disaster bearing body in the study area,the vulnerability distribution of the disaster bearing body in the study area is obtained.According to the calculation model of landslide risk evaluation,the current distribution of landslide risk level in Tongjiang County is calculated.Through the design experiment,the following three conclusions are obtained:(1)Taking decision tree,neural network and SVM as basic learners,using regular grid and slope unit as evaluation unit respectively,a comparative experiment is constructed.Through experimental verification,it is found that when grid unit is used as evaluation unit and neural network is used as basic learner,the prediction accuracy of landslide prone evaluation model is the highest.In this model,the AUC of 2013 and 2015 were 0.770 and 0.731 respectively when the area under ROC curve was used as the verification index.(2)Based on the CAMarkov model,this thesis discusses the influence of the number of iterations and neighbor types of Ca Markov model on the prediction accuracy.According to kappa coefficient as consistency detection index,when the number of iterations is 47 and neighbor type is 3*3von Neumann,it has high consistency(kappa coefficient 0.8121),so the model can predict and evolve the landslide disaster prone performance in the study area.(3)Taking the weighted density results of the disaster bearing body as the value index of the research unit,and combining the prediction results of the prone space in 2019 obtained by Ca Markov model,the risk zoning of Tongjiang County landslide can be obtained by grid superposition operation and grading the results.The research results will be taken as the basis for future development and construction as well as disaster prevention.In order to reduce the occurrence of disasters and the losses caused by disasters,we should take measures to control the high-risk areas,such as building retaining walls,anti slide piles and other facilities.