| Wencheng County of Wenzhou City is located in the south of Zhejiang Province.There are many hilly areas in this area,with steep terrain and complex geological development.Local rock mass in the area is relatively broken,and the slope stability is weakened under the influence of heavy typhoon rainfall and bad construction.In recent years,many landslide disasters occurred in Wencheng County,which caused serious damage to residents’ life and property safety.Therefore,it is of great significance to make an in-depth study of landslide disaster in Wencheng County and identify the areas with high spatial probability of landslide.The prediction of landslide hazard susceptibility can provide practical and reliable theoretical basis for the identification of potential landslides in the region.It is an effective means of disaster prevention and mitigation to monitor the landslide prone areas.However,the accuracy of landslide susceptibility prediction is affected by multiple factors,such as whether the landslide catalog information is accurate,whether the classification method and proportion of training set and test set are reasonable,whether there are errors in the establishment strategy of landslide data set,and whether the machine learning model type and parameter setting are appropriate.These factors affect the accuracy of landslide potential probability and the reliability of disaster prevention and mitigation measures.Therefore,how to reduce the uncertainty in the process of landslide susceptibility prediction and provide more scientific and reasonable landslide disaster prevention and control means is a difficult and hot issue at the present stage.In theory,landslide susceptibility prediction is to use landslides in the past to predict the potential area of landslide occurrence in the future,and its modeling has significant time sequence characteristics rather than simple spatial randomness characteristics.However,the current research on regional landslide susceptibility at home and abroad mainly focuses on the spatial scale characteristics of the catalogued data,while the temporal scale characteristics are only applicable to the physical simulation of the evolution mechanism of individual landslides.The temporal effect of regional landslide susceptibility is often ignored.For example,the training test set of the model is divided randomly from the perspective of spatial distribution.This may lead to the use of late landslides to predict the situation of early landslides,which is contrary to the prediction theory of landslide susceptibility.In addition,most susceptibility studies use all landslide catalogued data for susceptibility modeling,and the more landslide data the better.This leads to a significant reduction in the accuracy of the sample data set,and some "old slippery slope" of catalog errors is also involved in the susceptibility evaluation process.Landslide data on a large time scale involved in modeling will magnify the accumulation of uncertainty in research results.In order to solve the above problems,this paper carried out a landslide susceptibility prediction study based on Wencheng County.The relevant contents and research results are as follows:(1)Study and analyze the basic geological conditions of Wencheng County,obtain the temporal and spatial information and basic geographic data of existing landslides in the county,analyze the characteristics of landslides in Wencheng County and refer to the selection of environmental factors in the literature of landslide research in Wencheng County.Eleven basic environmental factors,such as landform,geological structure,hydroclimate,vegetation cover and surface building,which affect landslide development,were selected as the evaluation indexes of landslide susceptibility.The correlation and nonlinear correlation between landslides and various environmental factors are analyzed.(2)The landslide-non-landslide samples connected with environmental factors were divided into two different training and test sets according to the principle of landslide time sequence and spatial randomness.The partition ratio was set as 9:1,8:2,7:3,6:4 and 5:5 respectively to avoid different proportions affecting the research results.Finally,support vector machine(SVM),multilayer perceptron(MLP)and random forest(RF)models were used to predict landslide susceptibility and ROC accuracy and distribution law of susceptibility index were used to carry out modeling uncertainty analysis.It can be seen that the classification of time sequence conditions has higher prediction accuracy and lower uncertainty on the whole,and this classification method can make the prediction results of landslide susceptibility more reliable.(3)The existing 128 landslide catalogs were screened and updated,and the T landslide samples that were not screened and updated were constructed respectively,and the landslide data sets of T1,T2 and T3 in time increment condition after screening and updating were selected with 1990 as the time starting point and 2005,2015 and 2022 as time nodes.The landslide susceptibility model was established by SVM,MLP and RF machine learning models.Finally,the uncertainty of susceptibility and the influence of different machine learning models on the relative importance of environmental factors are analyzed.The results show that the prediction accuracy of susceptibility in T3 condition is the highest and the uncertainty is the least.The worst conditions were all the T conditions that were not screened and updated.It can be seen that the establishment method of data set considering time effect is necessary for the study of landslide susceptibility.In addition,the unfiltered updated data set interferes with the mainstream machine learning model,leading to a large error in the identification of the importance of environmental factors.The study also found that the relative importance of environmental factors will also change with the change of time increment condition,but the change range of RF model is much smaller than that of the other two models. |
PDF Full Text Request