Landslide Susceptibility Assessment Considering Time-Variance Of Dynamic Factors

Landslide is a common adverse geological phenomenon,which has the characteristics of wide distribution,high frequency and strong destructiveness,and seriously threatens the national economy and people’s safety.Landslide susceptibility assessment is the basic work of landslide prevention and control.According to the landslide survey data and geological environment conditions in the study area,the influence of the combination characteristics of various disaster-causing factors in the landslide disaster-pregnant environment on the occurrence of landslides is analyzed.Based on GIS,the study area is divided into different levels of susceptible areas,which provides the scientific basis for the implementation of landslide prevention and control policies.The existing study of landslide susceptibility assessment focuses on the spatial effect caused by the heterogeneity of the landslide itself and its static disaster-causing factors,without ignoring the time-varying of dynamic factors such as land use,normalized difference vegetation index(NDVI)and population density in the disaster-pregnant environment,which reduces the accuracy of the assessment results.Boshan District of Zibo City,Shandong Province was taked as the study area,the landslide data and geological environment conditions in Boshan District were investigated,and the landslide disaster-causing factors were extracted based on Arc GIS10.2 and ENVI5.3.The disaster-causing factors were conducts correlation analysis and collinearity test on.The tested disaster-causing factors were divided into static disaster-causing factors and dynamic disaster-causing factors.Three assessment factor combinations of static disaster-causing factor+dynamic disaster-causing factor 2021 measured value,static disaster-causing factor+dynamic disaster-causing factor each year measured value,static disaster-causing factor+dynamic disaster-causing factor interannual variation value were established.Three assessment factor combinations were input into five machine learning models(random forest model,logistic regression model,support vector machine model,stacking integrated model and convolutional neural network model)respectively.The prediction accuracy and assessment results of different factor combinations and different models were compared.The differentiation and factor detection function of geographic detector was used to analyze the interpretation degree of land use to landslide susceptibility in Boshan District.The interaction detection function was used to analyze the interaction of land use change to landslide susceptibility.Based on the Arc GIS10.2,the relationship between the time-variance of dynamic factors such as land use,NDVI and population density and the spatial distribution of landslide susceptibility was analyzed.The main conclusions are as follows:(1)Among the three assessment factor combinations,the assessment factor combination3 of static disaster-causing factor+dynamic disaster-causing factor annual change value is the most reasonable.Compared with the AUC values of the model under factor combinations 1(static disaster-causing factor+dynamic disaster-causing factor measured value in 2021)and2(static disaster-causing factor+dynamic disaster-causing factor measured value in each year),the AUC values are increased by 0.0546 and 0.0310 on average,and the verification accuracy is increased by 0.0251 and 0.0103 on average.Among the five machine learning models,the convolutional neural network model has the best prediction performance.Compared with the random forest model,logistic regression model,support vector machine model and stacking integrated model,the AUC value of the convolutional neural network model is increased by0.0470,0.0423,0.0267 and 0.0107 on average,and the verification accuracy is increased by0.0454,0.0390,0.0408 and 0.0050 on average.Among them,the landslide susceptibility assessment of the convolutional neural network model under factor combination 3 is the most reasonable,the AUC value is 0.92,and the verification accuracy is 0.9418.(2)Taking the convolutional neural network model under factor combination 3 as the benchmark model,the landslide susceptibility assessment results of different models and different assessment factor combinations in Boshan District were compared.The model under factor combination 1 and 2 overestimated the effect of elevation and underestimated the effect of distance from river.Other models such as stacking integration overestimate the effect of distance from fault and underestimate the effect of distance from road.The assessment factor combinations has a great influence on the assessment results of landslide susceptibility in Boshan District.The assessment results of landslide susceptibility based on the model of factor combination 1 and 2 have a strong tendency of extreme classification,which is easy to produce overestimation and underestimation.(3)The land use structure and spatial distribution of Boshan District are optimized year by year.The proportion of extremely high susceptible areas in land use change areas is large,especially in land use change areas such as bare land→forest land,cultivated land→artificial land,garden land →water area,forest land → water area and water area →forest land,the probability of landslide susceptibility is large.The proportion of excessive and moderate population mobility areas is relatively small,mostly distributed near Boshan urban area.The proportion of extremely high susceptible areas in excessive population mobility areas is the largest,and the proportion of extremely susceptible high areas in stable population mobility areas is the smallest.The area with severe interannual variation of NDVI accounts for a large proportion of extremely susceptible areas,and the area with stable NDVI changes accounts for a large proportion of extremely low susceptible areas.In the process of engineering activities and afforestation,attention should be paid to the schedule to avoid large-scale and short-term factor changes.