Landslide Susceptibility Assessment Based On 3S Technology And Multi-model Fusion

With the increasing exploitation of groundwater and various mineral resources,the stress of rock and soil structure and the balance of groundwater resources are changed,and the influence of extreme weather leads to the occurrence of landslides frequently.The evaluation results of landslide susceptibility can be used as basic data for scientific disaster prevention and control.By judging the high-prone areas,the prevention and treatment of key areas in advance can make disaster prevention more scientific and reasonable.At present,there are some problems in the evaluation of landslide susceptibility,such as insufficient correlation test of factors,low accuracy of evaluation model,and whether the integration of multiple models can improve the accuracy.In this paper,relevant studies were carried out from two aspects of vulnerability factor correlation test and improvement of vulnerability model,and the following results were obtained:(1)Select multi-source data to make landslide susceptibility factors,and build a database of candidate factors.Taking Yibin City of Sichuan Province as the research area,multi-source data including geology,remote sensing and meteorology were selected.Choose digital elevation model(DEM),slope,slope direction,degree of relief,slope type,section curvature,curvature,point of Interest(POI)Density,the ground motion peak acceleration(PGA),is apart from the fault distance,the distance from the river,away from the road distance,normalized difference vegetation index(NDVI),precipitation,strata lithology,prone to a total of 16 for evaluation Child,build the candidate database.(2)Analyze the positional relationship between landslide disasters and susceptibility factors in Yibin City.A spatial analysis of the location relationship between landslide disaster points and 16 susceptibility influencing factors in Yibin City reveals that most landslides in Yibin City are distributed in the western,central and southern regions;near rivers and steep slopes,landslides are more concentrated;precipitation There are more landslides in high-volume areas;the distribution pattern of landslides in the southeast of Yibin City is basically distributed along the Huayingshan fault zone.(3)Various methods were used to test the correlation of landslide impact factors,and a vulnerability evaluation factor database was built.Each method for correlation test of landslide impact factors has its own advantages and disadvantages,and only one method for correlation test of factors is insufficient,and sometimes the factors with correlation may be omitted.Therefore,in this paper,three methods of multicollinearity test,stepwise discrimination and Pearson product moment correlation coefficient are selected to test the factor correlation.The multicollinearity test has a high tolerance to factors,and all the factors in the candidate library meet the conditions;the stepwise discriminant method can better screen the factor correlation,but the covariance matrix used by this method is sometimes unstable,and the correlation is relatively low.Weak factors cannot be screened well;the Pearson product-moment correlation coefficient can be mutually verified with the stepwise discriminant method,and the choice of factor correlation coefficient can be adjusted according to the actual situation,which is more flexible.The multicollinearity test found that all the 16 factors met the conditions,and the stepwise discriminant analysis found that 10 factors met the conditions.The Pearson product moment correlation coefficient found that the slope had a strong correlation with the relief of the terrain.After the topographic relief was eliminated,this paper used nine factors to be selected,namely,the slippery rock group,slope,peak acceleration of ground motion,DEM,NDVI,distance from river,distance from fault zone,precipitation and distance from road,to participate in the construction of the evaluation model.This paper found that the multiple methods are more stringent than the single method to test the correlation of factors,which can ensure that the selected evaluation factors meet the condition of no correlation.(4)Integrate multiple evaluation models to evaluate the susceptibility of landslides in Yibin City.Use frequency ratio(FR)method,Logistic(LR)model,information volume(I)method,radial basis(RBF)neural network,multi-layer(MLP)neural network 5 single models to evaluate the landslide susceptibility of Yibin City.Then,for the problem that the information quantity model does not consider the factor weight,use Logistic model,RBF neural network,MLP neural network and information model to integrate,get LR-I model,RBF-I model,MLP-I model,LR-RBF-I model,LR-MLP-I model,RBF-MLP-I model,LR-RBF-MLP-I model,a total of 7 integrated models.(5)Receiver operating characteristic curve(ROC)was used to evaluate the accuracy of 12 landslide susceptibility zones in Yibin City.The area(AUC value)between receiver operating characteristic curve(ROC)and coordinate axis was calculated,and the accuracy of FR model was 75.80%.The accuracy of MLP model was 74.80%.The accuracy of LR model was 76.90%.I model precision is 76.30%;The accuracy of RBF model was 74.00%.The accuracy of RBF-I model was 76.00%.The accuracy of MLP-I model was 75.40%.The accuracy of LR-I model was 77.10%.The accuracy of LR-RBF-I model was 76.40%.The accuracy of LR-MLP-I model was 76.20%.The accuracy of RBF-MLP-I model was 76.40%.The accuracy of LR-RBF-MLP-I model was 76.20%.This shows that it is not that the more models that participate in the integration,the higher the accuracy.The reason for the analysis is that the weight allocation of each factor is not consistent among the models involved in weighting.Excessive model integration will make the weight of factors with large differences tend to be the same,resulting in lower accuracy.(6)Subjective assignment method was used to evaluate the accuracy of 12 landslide susceptibility zones in Yibin City.Subjective assignment method(classification of the landslide point in the low-risk area was assigned 0.2 points;The classification of landslide points in the middle prone area was assigned 0.8 points.The classification of the landslide points in the highly vulnerable areas is assigned 1 point).The results of 12 models show that: The accuracy of FR model was 82.46%.The accuracy of MLP model was 81.64%.The accuracy of LR model was 81.87%.I model accuracy is 86.90%;The accuracy of RBF model was 71.46%.The accuracy of MLP-I model was 85.50%.The accuracy of LR-I model was 88.30%.The accuracy of RBF-I model was 84.09%.The accuracy of LR-RBF-I model was 85.96%.The accuracy of LR-MLP-I model was 86.67%.The accuracy of RBF-MLP-I model was86.67%.The accuracy of LR-RBF-MLP-I model was 87.13%.The LR-I model has the highest accuracy,indicating that the integrated model can indeed improve the accuracy of landslide susceptibility zoning.To sum up,this paper built Yibin vulnerability impact factor database.In view of the problems that evaluation factors are correlated in previous work,three methods were used to screen the factors of vulnerable areas,and a total of 9 unrelated evaluation factors were screened out.Aiming at the problem of low precision of single model,multiple models were integrated to explore whether the integrated model could improve the evaluation accuracy of vulnerability zoning.Finally,two methods of ROC curve and subjective assignment are used to verify the accuracy of the results of12 susceptibility partitions.Finally,it is found that the accuracy of the LR-I model after the integration of the Logistic model and the information model is the best.The conclusions and research ideas of this article have good reference and reference value for future susceptibility or risk zoning evaluation.