Considering Landslide Types Susceptibility Prediction Modeling And Detailed Hazard Warning

Landslides are seriously threatening people’s lives and property,and causing serious damage to natural resources,ecosystems and infrastructure.By closely integrating geographic information system(GIS)technology with machine learning models,more efficient and accurate susceptibility prediction models can be constructed in the form of images and numbers.In general,there are several important issues in landslide susceptibility prediction modeling,the following analyzes these issues separately:The calculation of non-linear correlation between landslide inventory and its environmental factors and different machine learning models are important factors that affect the uncertainty of landslide susceptibility predictive(LSP)modeling.In order to study the changing patterns of LSP under the influence of these uncertain factors,taking Huichang County,China as an example,417 landslides and 12 environmental factors are obtained.Four machine learning models,including Logistic Regression,Bayesian Networks,Multi-layered Perceptron,Support Vector Machines,and C5.0 Decision Trees,are respectively coupled through five connection methods such as Probability Statistics,Frequency Ratio,Information Volume,Index of Entropy,and Weight of Evidence,forming 20 types of work to carry out LSP.It is compared with the individual LR,BN,SVM and C5.0 DT models with the original environmental factor data as input variables for exploring the LSP modeling pattern.Most of the existing landslide susceptibility prediction modeling is only based on a certain type of landslide,without considering the connections and differences between multiple types of landslides.Aiming at this problem,considering landslide types susceptibility prediction model is proposed and its modeling uncertainty is analyzed.Taking Huichang County as an example,three susceptibility models considering landslide types are carried out:(a)The united method is directly combining different types of landslides,and then carrying out the susceptibility modeling.(b)Probability method is calculating each type landslide susceptibility index through the probability and statistical formula.(c)Maximum susceptibility method is selecting the larger landslide susceptibility index as the final landslide susceptibility value.Then the uncertainty of the landslide susceptibility results considering landslide type is evaluated.The critical rainfall threshold usually refers to the critical rainfall value when the landslide is unstable and damaged by a certain rainfall process,and is often used to distinguish the rainfall-induced landslide(unstable state)and non-landslide(stable state).On the basis of the landslide susceptibility mapping considering the types of landslides,the risk warning modeling research of rainfall-induced landslides are further considered.The main research contents and results are as follows:(1)The machine learning model based on WOE has the least uncertainty in landslide susceptibility prediction.It can be seen that WOE has better nonlinear correlation performance than other methods.The susceptibility accuracy predicted by the single machine learning model is slightly lower than that of the coupled model,but its modeling efficiency is higher than that of the coupled model.In reflecting the spatial correlation between landslides and their environmental factors,the coupled model considering the connection method has significant advantages.(2)Among the five machine learning models selected in this paper,the C5.0 DT and SVM models have better prediction performance,followed by BN,MLP and LR models respectively.In landslide susceptibility modeling,machine learning model factors are more sensitive than connection method factors.(3)The overall accuracy of the predicted results of the three landslide susceptibility models based on SVM and C5.0 DT considering the types of landslides is good,and the susceptibility results have high similarity and little difference.The uncertainty of predicting landslide susceptibility by the united method is lower than that of the probability method and maximum susceptibility method,but the performance of various landslide susceptibility prediction models considering different landslide types needs to be verified by more cases.(4)Two landslides in 2019 are used for verification and he results show that the two landslides have good early warning effect in traditional hazard and full-probability hazard.Both landslides fell in the extremely high susceptibility area(above 0.8),as well as in the rainfall threshold level 5 special warning area(90%～100%)and extremely high risk(above 0.7)area.