Landslide Susceptibility Mapping In Jiuzhaigou Earthquake Region Based On Deep Learning Models

China is characterized by complex and diverse terrain,vast mountainous areas and strong geological structure development,resulting in a very wide distribution of geological hazard potential.Landslide geological disaster is one of the most typical geological hazards in China,and in recent years,China’s geological hazards in general are still in a multi-prevalent situation.According to the statistics of national geological disaster notification and other related information,the number of landslides occurring between 2010 and 2021 accounts for 70% of the total number of geological disasters,with the characteristics of wide impact,large number of occurrences,strong destructive force and extremely serious damage caused,especially landslides triggered by strong earthquakes have brought adverse effects to the development of southwest China.In this context,it is of great practical significance to carry out the study of earthquake landslide susceptibility assessment and to quantitatively analyze the results of landslide susceptibility assessment in earthquake-affected areas for post-disaster emergency response,disaster prevention and mitigation,post-disaster reconstruction,and for future medium-and long-term planning of the affected areas.Landslide susceptibility assessment is an important part of geological hazard risk assessment.This thesis takes earthquake landslide susceptibility assessment as the selected topic,takes deep learning model as the technical guide,uses geographic information data and remote sensing images as the data source,addresses the problems of constructing an earthquake landslide susceptibility assessment process based on deep learning model,the existence of hyperparameter optimization and the inadequacy of sample data set sampling,conducts research on earthquake landslide susceptibility assessment based on deep learning,and carries out application verification in the hardest hit areas of Jiuzhaigou earthquake.(1)The process of landslide susceptibility assessment of Jiuzhaigou earthquake based on deep learning model is established.With the principle of geological hazard risk assessment as the theoretical guide,the deep learning algorithm model is applied to the landslide susceptibility assessment process from five aspects,namely,the establishment principle and construction process of the landslide susceptibility assessment sample set,the processing and preference of landslide susceptibility assessment impact factors,the application and improvement of the model,the evaluation and comparison of the accuracy of the landslide susceptibility assessment model,and the analysis of the validity,scientificity and rationality of the specific landslide susceptibility assessment results.The key steps in the process of landslide susceptibility assessment are explained and sorted out.The application demonstration is carried out in conjunction with the specific case of Jiuzhaigou earthquake landslide,and the sample data set of landslide susceptibility assessment is collected through regional landslide survey and visual interpretation of remote sensing images;9influencing factors are selected from 16 influencing factors for model construction through feature correlation and importance analysis;landslide susceptibility assessment in the study area is carried out through specific model application and accuracy evaluation.(2)A landslide susceptibility assessment model based on Ant Colony Optimization(ACO)and Deep Belief Networks(DBN)is proposed.In order to address the problem of insufficient hyperparameter optimization of deep learning models in landslide susceptibility assessment research,the study compares one deep learning model,Deep Belief Network,two statistical machine learning(ML)models:Support Vector Machine(SVM),Random Forest(RF),and the integrated models ACO-RF,ACO-SVM,and ACO-DBN,which are hyperparametric optimization of these three models by ant colony optimization algorithm,respectively.Through the model modeling and hyperparameter optimization algorithm improvement design,the research results show that the DBN as a deep learning model has an accuracy rate of93.11% after optimization,and its accuracy performance is better than the optimized statistical machine learning models RF and SVM,and its landslide density distribution is more reasonable.The three integrated models outperformed the unoptimized original model in terms of each accuracy metric and the receiver operating characteristic(ROC)curve,and their optimization results were statistically significant.This demonstrates the potential of using deep learning models with hyperparameter optimization in landslide susceptibility assessment research.(3)A landslide susceptibility assessment method based on Gaussian heat map sampling and Convolutional Neural Network(CNN)is developed.Gaussian heat map sampling is used to enrich the variety of landslide sample datasets at the input of the deep learning model to improve the accuracy and precision of the landslide susceptibility assessment results.The sampling method constructs a landslide susceptibility Gaussian heat map neural network model by combining convolutional neural networks.After the optimization of model hyperparameters,by comparing Gaussian thermogram sampling with the traditional positive and negative sample sampling methods,the accuracy of the assessment results(95.30%)and the F1 value of the comprehensive assessment index(95.13%)of the Gaussian thermogram sampling method improved by about 2% compared with the traditional method,and the sampling method also effectively improved the accuracy of the probability distribution map of landslide susceptibility.Through the sensitivity analysis of Gaussian thermogram scale sampling,it is found that the mesoscale Gaussian thermogram is more suitable for the application of landslide susceptibility assessment in Jiuzhaigou earthquake area,as well as the correctness of the assessment results is verified by the future occurrence of landslides.The results demonstrate that the Gaussian heat map sampling method can enrich the variety of landslide sample datasets,and the method enables deep learning to improve the precision as well as accuracy in landslide susceptibility assessment results.