Evaluation Of Landslide Susceptibility In YANBIAN County Based On Multi-model Comparison

In recent years,affected by the "5.12" Wenchuan Earthquake,"8.30" Panxi Earthquake,"6.17" Changning Earthquake,as well as the frequent extreme rainfall climate and unreasonable human activities,the geological environment in the mountainous regions of Sichuan Province has become more fragile,causing a large number of new geological hazards such as new landslides,collapses,and debris flows.They are particularly prominent in the mountainous areas of Sichuan with large terrain fluctuations and complex geological structures.Yanbian County is located at the southernmost tip of Sichuan.It is dominated by mid-high mountain tectonic erosive landforms.The geological conditions are complex and the rainy season is concentrated,which causes frequent geological disasters such as landslides in the region,which seriously threatens local people 's life and property safety and hinders economic stability,Healthy and sustainable development in the region.Carrying out landslide susceptibility evaluation studies and analyzing the probability of landslides in different regions from a macroscopic perspective can provide scientific references for local governments to make decisions on disaster prevention and mitigation,and it is also of great significance to people's life and property safety and local socio-economic development.This paper introduces rough set,logistic regression and machine learning model theory,and combines "3S" technology,MATLAB programming and other technical methods to study the effects of the logistic regression model,support vector machine model and BP neural network model in the evaluation of landslide susceptibility,using Yanbian County landslide as the study object.Based on the geology and geomorphology of the study area and the detailed survey data of existing landslides,the in-depth analysis of landslide development characteristics,factor sensitivity,and zonation of landslide susceptibility were carried out,and the accuracy of the evaluation results of each model was compared and analyzed to obtain optimal landslide susceptibility evaluation model for Yanbian County.In the course of systematic research work,this paper has obtained the following research results and understandings:(1)Identified the spatial distribution pattern of landslides in the study area and carried out the analysis of landslide development rules.Yanbian landslide disaster points in the distribution of the basin has the characteristics of relative concentration.The number of landslide disaster points in the Ganyu River Basin is the largest,reaching 45.56%.Most of the landslides are mainly shallow loose soil landslides,and the scale of landslides is small.Moreover,they occur mostly during the rainy season and flood season.(2)Established grading criteria for landslide evaluation factors.In order to ensure that the factor classification can preserve the spatial attribute characteristics of the landslide to the greatest extent,the paper first extracts the attribute values of the 14 impact factors that are cataloged in the field,and finally establishes a factor classification standard based on the frequency distribution characteristics of the attribute values.(3)Analysis of the sensitivity of the impact factors.The magnitude of landslide sensitivity for each classification of landslide impact factors was analyzed by using the coefficient of determination(CF)method.The mechanisms by which each impact factor induces landslides and the causes of high landslide sensitivity intervals are summarized.(4)Screening evaluation factors.The pre-selected 14 impact factors were screened using Rough Set Theory(RS),removing the 4 factors of redundant or unimportant terrain undulation,LS coefficient,Melton intensity,and rainfall during the rainy season,and retaining the elevation,slope,slope length,engineering rock group,slope structure,fault distance,water distance,land use,and road distance 10 evaluation factors.(5)Construction of landslide susceptibility evaluation model.The CF values of 10 factors were extracted into the training samples as input values of the model,and three models of logistic regression,support vector machine,and BP neural network were established for training respectively.The spatial distribution trend of the landslide susceptibility predicted by the three models is generally consistent.The high susceptibility areas are mainly distributed in the Ganyu River and its tributaries along the nearby valleys in the middle,which are mainly weathered and fragmented metamorphic rocks,and in the Aning and Hongguo River basins,which are dominated by loose mudstone interlays in the Xigeda Formation in the south;The low susceptibility areas are mainly distributed in the high-altitude Gesala township,which is dominated by carbonate rocks.(6)Verification of the accuracy of landslide susceptibility.The zoning results were verified using two indicators,the relative density of the landslide point LRD and the area under the ROC curve,AUC.From the LRD index,the LRD value of the BP model > SVM model > LR model in the same subregion,the difference is particularly prominent between the extremely low-prone areas and extremely high-prone areas.From the AUC index,the AUC of the LR model is 0.7739,the AVM of the SVM model is 0.8109,and the AUC of the BP model is 0.8632.Combining the two indicators shows that the BP model is better than the SVM model,and the SVM model is better than the LR model.