Uncertainty Analysis Of Rainfall-induced Landslide Susceptibility Prediction And Risk Assessment Modeling

As a kind of sudden disaster,landslides occur frequently all over the world,seriously hindering economic and social development and endangering people’s life and health.As one of the countries most seriously affected by geological disasters,the vast mountainous areas of China suffer huge losses due to landslide disasters,so it is of great significance to effectively prevent and control landslide disasters.In landslide disaster research,landslide susceptibility prediction can obtain the spatial probability of landslide occurrence,and rainfall threshold model can predict the specific time of landslide occurrence according to the time probability.Therefore,landslide risk assessment based on susceptibility prediction and rainfall threshold calculation can provide scientific guidance for prediction of spatial and temporal probability of regional rainfall-type landslide,and effectively promote landslide warning and prevention work.However,there are many uncertainties in the process of landslide susceptibility modeling,such as the selection of conditioning factors and model construction.The traditional critical rainfall threshold can only qualitatively divide the rainfall threshold level,and there are uncertainties with low spatial identification.In addition,the highway density in the study area is considered as the static factor inducing landslide,which can more objectively verify the inducing factors affecting the spatio-temporal probability of landslide to some extent,and provide a reference for the uncertainty analysis of rainfall threshold model prediction.Therefore,in view of the above problems in the vulnerability prediction and rainfall threshold of landslide risk assessment,this paper takes Anyuan County of Jiangxi Province as the research object to carry out regional landslide vulnerability prediction and risk assessment.The main research contents and results are as follows:(1)Explore the effects of different selection methods of conditioning factors and their optimal combinations on uncertainty of Landslide Susceptibility Prediction(LSP).Firstly,431 landslide locations and four types of 29 conditioning factors were obtained.Correlation Analysis(CA),Linear Regression(LR),Principal Component Analysis(PCA),Rough Set,RS and Artificial Neural Network(ANN)were used to select the optimal combination of conditioning factors from 29 conditioning factors.The results show that the conditioning factor selection method cannot guarantee the performance improvement of LSPS.The LSPS based on LR or ANN are generally more accurate than the LSPS based on RS,CA or PCA,and the comprehensive model without factor selection is generally better than the LSPS based on conditioning factor selection.(2)The results of conditioning factor selection were further combined with the results of Multilayer Perceptron(MLP),Support Vector Machine(SVM),SVM),ChiSquare Automatic Interaction Detection(CHAID),Random Forest(RF)and four machine learning models were combined to form 20 combination models such as CAMLP and CA-SVM.In addition,comprehensive MLP,SVM,CHAID and RF models were constructed to compare all conditioning factors as input variables.Among the four kinds of machine learning,the coupling model based on RF model has the best prediction accuracy,which is higher than the prediction accuracy of CHAID,SVM and MLP.(3)Taking Anyuan County of Jiangxi Province as the study area,104 rainfall-type landslide events from 1986 to 2004 were selected as the basic data,and Early Effective Rainfall(EE)and Duration of Rainfall(Duration of Rainfall)were used.D)The e E-D threshold method is used to calculate the time probability(P)level of landslide induced by different rainfall;Then,the logistic regression model was used to fit the nonlinear mapping relationship between P and EE and D corresponding to each critical rainfall threshold curve,and all the detailed probability(total probability)critical rainfall thresholds of rainfall-induced landslide were obtained.(4)Road density was introduced,road density ratio(R)was used as static induction factor,rainfall was used as dynamic induction factor,and the total probability static and dynamic rainfall threshold was obtained.The results show that there are 18 landslides in the traditional and static and dynamic critical rainfall thresholds,while there are 14 and 12 landslides in the total probability critical rainfall threshold and the total probability static and dynamic rainfall threshold,respectively,which are above60%.(5)The landslide susceptibility and rainfall threshold results were coupled to evaluate the total probability risk and static and dynamic total probability risk results,which were verified by two landslides.The results show that the two rainfall-type landslides have good warning effects in the two types of risk assessment,and there is little difference between them.One landslide falls in the special warning area of grade5 of rainfall threshold(90%～100%),and the other one falls in the grade 4 warning area(60%～90%),and the risk is greater than 0.8.