Effects Of Landslide Inventory Error And Susceptibility Classification On Landslide Susceptibility Evaluation Modeling

China has a vast land area,and mountainous areas account for 70% of the land area.However,landslides are more likely to occur in mountainous areas,so many areas of China are threatened by landslides.There are many mountainous areas in Shangyou County of Jiangxi Province in China.Due to heavy seasonal rainfall,construction and development activities and other human engineering activities,many landslides have occurred in recent years,resulting in the destruction of related construction sites and roads,and casualties.Therefore,Shangyou County is the key area of landslide prevention and control in Jiangxi Province.Landslides cause a large number of life and property losses all over the world every year,and how to predict the spatial probability distribution of landslides has become the focus of researchers all over the world.Landslide susceptibility evaluation modeling as a research scheme can accurately predict the spatial probability of potential landslides in a certain area to a certain extent.In the existing relevant studies,the evaluation of the susceptibility of geological disasters such as landslides mainly obtains the potential high susceptibility of landslides in the study area by exploring the internal and external factors of landslides,such as geology,soil,vegetation,climate and human engineering activities;Or by improving or creating a new machine learning model to improve the prediction accuracy of landslide susceptibility,so as to improve the evaluation effect.Few researchers pay attention to the relevant uncertainty in landslide susceptibility modeling.The basic data related to landslide inventory is a crucial basic step for landslide susceptibility evaluation.In the study,the shape characteristics and position of landslide inventory affect the boundary and internal area of landslide,and the information of environmental factors contained in each landslide also produces some errors.At the same time,the classification expression of susceptibility is the last link of landslide hazard susceptibility evaluation,which is at the key point of the whole process.The selection of classification method of susceptibility affects the characteristics of the final landslide susceptibility map to a large extent.The key performance of an effective landslide susceptibility map is that it has a smaller area of high susceptibility and can identify more historical landslides,which is also the main basis for future decision making.Therefore,in view of the relevant problems in the above research,this paper takes Shangyou County of Jiangxi Province as the research area to carry out the study on the influence of landslide inventory error and susceptibility classification on landslide susceptibility modeling,and explores the use of semi-supervised and other ideas to overcome the impact of landslide inventory error.the research object to study the influence of landslide inventory error and susceptibility classification method on landslide susceptibility modeling.The main research contents and results are as follows:(1)The landslide inventory information of shangyou county from 1970 to 2003 was collected and 337 landslides were obtained.The characteristic expression of landslide was mainly the polygonal surface of landslide and the central point of complete landslide area.In this paper,a circular landslide boundary obtained from the center of landslide is used as the third landslide shape feature expression method.The three different boundary and spatial shapes are used as the expression methods of landslide shape characteristics for landslide susceptibility evaluation.(2)Ten kinds of environmental factors including disaster pregnant factors and inducing factors are selected as the model evaluation factors of the study area.These factors are obtained through historical data,land information and a large number of documents.Then,the landslide shape and its correlation with environmental factors are established based on landslide point,buffer circle and accurately obtained polygon;Then,multi-layer perceptron(MLP)and random forest(RF)are selected to build six evaluation models based on point,circle and polygon.Finally,ROC accuracy and distribution law of susceptibility index are used to carry out modeling uncertainty analysis.It can be seen that taking landslide points or buffer circles as landslide shape features will increase the uncertainty of susceptibility results,and using accurate polygon surface to express landslide shape features can obviously more effectively ensure the accuracy and reliability of susceptibility results.(3)Select the best group of working conditions,that is,taking the landslide polygon as the expression of landslide shape characteristics and evaluating the landslide susceptibility of Shangyou County Based on RF model as a case,and put forward the quantitative classification standard of landslide susceptibility.The susceptibility map,area ratio and the classification accuracy of susceptibility based on frequency ratio were used to verify.It can be seen that when evaluating landslide susceptibility,taking natural breakpoint method and K-means clustering method as landslide susceptibility classification method can make the result more scientific and reasonable.(4)On the basis of the previous studies,the polygon surface of landslide was used as the landslide expression feature,and the polygon surface was randomly shifted by30 m,50 m,70 m,90 m and 110 m to form grade errors.Also build original data-based,30 m...110 m error-based MLP and RF models were used to predict the landslide susceptibility index.The spatial position errors were randomly simulated twice to avoid the randomness of modeling.Finally,uncertainty analysis was carried out.It can be seen that larger spatial position error of landslide will increase modeling uncertainty,but the error data can also be used within the range of 70 m.Therefore,it is feasible to use the existing spatial position data of landslide,which is difficult to avoid and has certain errors,for susceptibility modeling.(5)Semi-supervised machine learning model was used to evaluate the susceptibility of landslide under the original landslide condition to reduce the influence of spatial position error.It can be seen that the prediction accuracy of the original landslide susceptibility based on semi-supervised machine learning is significantly higher than that based on other spatial position errors.Therefore,the semi-supervised machine learning model is effective in reducing the impact of spatial position errors on landslide susceptibility evaluation modeling.