Machine Learning Based Landslide Susceptibility Evaluation And Slope Monitoring Method Optimization

China is located in a region with frequent crustal movements and large topographic relief,and is one of the countries with the most serious landslide disasters in the world.In order to reduce and prevent losses caused by landslide hazards,there is a need to evaluate the spatial probability of landslides for susceptibility and provide information guidance and direction.Ya’an is located at the intersection of the Qinghai-Tibet Plateau and Sichuan Basin,with complex topography and longitudinal fractures.In this thesis,Ya’an is selected as the study area,and landslide hazard impact factors in the study area are screened as susceptibility evaluation factors through a combination of fieldwork and literature research.Using 30 m raster cells as evaluation units,five types of landslide susceptibility evaluation models in the study area were constructed using statisticalbased frequency ratio model,weighted frequency ratio model and machine learningbased decision tree model,Light GBM model and Light GBM-frequency ratio model,respectively,and the models were compared in terms of rationality and evaluation accuracy,and the results of landslide susceptibility area classification were The optimal monitoring scheme was selected based on the results of landslide susceptibility area classification.The main research contents and the conclusions obtained are as follows:(1)Taking Ya’an City as the study area,the basic situation of the study area was analyzed in five aspects: geological conditions,natural environment,ecological environment,meteorology and hydrology and human engineering activities,and the spatial and temporal distribution patterns of landslides in the study area were summarized.The results show that spatially,landslides in the study area show the characteristics of distribution along water systems and roads,and are most dense in Hanyuan County,Yucheng District and Hanyuan County in the south-central part of the study area.Temporally,landslide disasters in the study area mostly occur from June to August when there is a large amount of rainfall.(2)According to the basic situation of the study area,18 landslide susceptibility influencing factors were initially selected from five aspects: geological conditions,natural environment,ecological environment,meteorology and hydrology,and human engineering activities.According to the calculation results,13 evaluation factors with greater weight,weaker correlation and no multicollinearity were retained as the evaluation index system of landslide susceptibility in the study area.Among them,geological conditions evaluation factors include the distance from thr fault,stratigraphic lithology,and peak acceleration of ground vibration;natural environment evaluation factors include curvature,slope direction,slope,topographic humidity index;ecological environment evaluation factors include vegetation type,normalized vegetation index;meteorological and hydrological evaluation factors include average annual rainfall,distance from water system;human engineering activities evaluation factors include distance from road,land use type.The spatial distribution relationship between each evaluation factor and landslide was summarized and analyzed.(3)Five types of landslide susceptibility evaluation models were established based on 30 m raster cells: statistical-based frequency ratio model,weighted frequency ratio model and machine learning-based decision tree model,Light GBM model,and Light GBM-frequency ratio model,respectively.The evaluation results were tested for reasonableness and accuracy by mathematical statistical method and ROC curve analysis method.The results showed that the evaluation results of the above five models were reasonable and showed high prediction accuracy.In the reasonableness test,the Light GBM-frequency ratio model has the smallest percentage of landslide point number and frequency ratio in the very low susceptibility area and the largest percentage in the very high susceptibility area,with the best zoning effect.In the accuracy test,the Light GBM-frequency ratio model also has the highest accuracy,so the model is the best landslide susceptibility evaluation model for the study area in this study.(4)The weights of key indicators for landslide monitoring were calculated by using the improved hierarchical analysis method,and three monitoring schemes were determined based on the evaluation results of landslide susceptibility in the study area.By comparing and analyzing the monitoring accuracy,timeliness,equipment maintenance cycle and economy of each scheme,the optimal monitoring scheme was selected for each level of landslide susceptibility zoning to provide data support and scientific basis for landslide warning and protection.