Landslide Susceptibility Assessment Based On Logistic Regression?Artificial Neural Networks And Random Forests Model In Lanzhou City

Lanzhou is one of the high incidence of geological hazards city,along with "the Belt and Road" and "China's Western Campaign" strategies,not only to speed up the construction and development of the city,but also exacerbated the associated human activity,but also led to Lanzhou landslide disasters more frequent,so it is necessary to analysis factors of landslide disaster and analysis landslide Sensitivity based on the mathematical and Machine learning models,on the one hand to facilitate the understanding of the distribution of Lanzhou landslide hazard and the formation of characteristics,on the other hand also provide the basis for decision making for the city Government to prevent the landslide.The method of the domestic regional landslide susceptibility assessment using the ability to solve nonlinear problems is weak,the accuracy of the evaluation by the limited accuracy of the sample data,and less comparative study between different models for assessing performance,and the results of the analysis of the main validation method is the accuracy of training to learn to measure,verify the method is relatively simple.In recent years with the development of computer technology and mathematical algorithms,the international development of a number of new evaluation methods,such as: SVM(Support Vector Machine)and RF(Random Forests),these new methods can quickly process large amounts of data,and for the characteristic data can be more accurately identified,so this study was to explore a new model for regional landslide susceptibility assessment of the suitability and sensitivity of partitioning ability to compare different methods for the future of the domestic regional landslide susceptibility assessment methods provide a basis for selection.The study is based on the factors of landslide hazards in the study area,selected seven landslide factors,the establishment of logistic regression,artificial neural network and random forest model combined with GIS-related tools to evaluate Lanzhou landslide hazard susceptibility,verification and analysis of the accuracy and applicability of different model results.Finally obtained the following conclusions:(1)Analysis of related factors of landslide in the study area,and ultimately selected elevation,NDVI,slope,aspect,lithology,topography and distance from the valley as influencing factors.After a random forest model analysis,different factors influence the degree of influence of the landslide,where topography,slope,DEM and NDVI To maximize the impact of the four factors,topography factors in loess and bedrock greatest impact on the landslide;slope the larger,the greater the probability of occurrence of landslides;elevation of 1800-2600 m range high landslide susceptibility;NDVI is within the range of 0.06 to 0.3,the higher the landslide susceptibility Understand and analyze the impact of the landslide-related factors contribute to a more thorough grasp of distribution and formation mechanism of landslide hazard.(2)Based on logistic regression,artificial neural network and random forest model for the study area for landslide hazard evaluation sensitivity three sensitivity evaluation shows moderate sensitivity study area zoning area were more than 28.9% of the total area,42.19% and 37.02% three evaluation results of the sensitivity distribution of the spatial distribution of existing landslide consistent sensitivity trends area are also consistent with the actual situation.The use of artificial neural network model and random forest in the study area high landslide susceptibility region landslide type of prediction,wherein loess landslides mainly in the study area declared Ravine-sulcus region,Gaolan Hill-muddy ditch Li Masha trench regions are distributed;bedrock landslides are mainly distributed in Ma Li Shagou region;composite landslides mainly in Gaolan Hill-mud ditch area,and the rest scattered distribution in the study area.Distribution predict landslide types and different types of existing landslide distribution is basically the same.(3)Respectively use the confusion matrix,ROC curve and the PS-In SAR results validate the results of the three models and contrast corrected percentage logistic regression accuracy rate of 86.20%;artificial neural network confusion matrix accuracy was 80.8%,ROC curve AUC was 0.8796;random forests confusion matrix accuracy rate of 82.90%,ROC curve AUC is 0.9118.PS-In SAR verify the accuracy rate of 62.01%,66.35% and 73.74%.Compare the results of three models,found that the result of random forests is more accurate,its evaluation results consistent with the existing distribution of landslide and concluded that random forest model is more suitable for the study area of landslide hazard Sensitivity Evaluation.