Study On Machine Learning Optimization Model And Interpretability Of Landslide Susceptibility

The mountainous environment in southwest China is widely distributed,and the geological environment and hydrological and climatic conditions are complex and changeable.It is the most serious landslide disaster area in China.Landslide has many disaster-causing factors and complex coupling mechanism,which makes it difficult to put forward effective prediction model.A large number of accumulated historical landslide and its impact factor data provide a good foundation for mining and analyzing the landslide susceptibility mapping model,and then put forward more effective disaster prevention ideas.However,the performance of the prediction model is usually affected by the evaluation unit and the evaluation factor,and the evaluation results are not persuasive enough because of the existence of the “black box model”.Therefore,it is of great significance to study the optimization of evaluation units,the optimization of factor screening methods and the construction of interpretable models to improve the accuracy and interpretability of the models for disaster prevention and control.In this paper,Chengkou County,Wuxi County,Wushan County,Yunyang County and Fengjie County in typical mountainous areas of northeast Chongqing are selected as the research areas,and the interpretable model of landslide susceptibility mapping based on unit optimization and factor optimization is studied from three aspects of evaluation unit optimization,factor screening method optimization and establishment of interpretable model.The main works and results of this paper are as follows.(1)By collecting the historical landslide data,3594 landslides occurred in the study area from 2000 to 2020 were counted,and the historical landslide catalogue in the study area was constructed.The temporal and spatial distribution characteristics and main inducing types of landslides in the study area were analyzed.Affected by the subtropical monsoon climate,the landslide in the study area is mainly between July and September,and rainfall is the main cause of landslide in the study area.(2)By summarizing five evaluation units commonly used in landslide susceptibility research,comparing the advantages and disadvantages of different evaluation units and applicable conditions,grid unit and slope unit are selected as the evaluation units used in this study.By setting different parameters,the r.slopeunits tool is used to divide 30 groups of different evaluation units.The results of evaluation unit optimization show that the evaluation results of SU16 and SU21 in slope unit are the best.(3)Based on the analysis of physical geography,geographical environment and landslide development characteristics in the study area,22 influencing factors under the influence of topography,geological conditions,environmental conditions and human engineering activities are selected to construct the landslide susceptibility evaluation index system.The geographical detector,recursive feature elimination and multicollinearity test were used to optimize the factor selection method.The optimization results showed that the model performance using recursive feature elimination method based on slope unit SU21 was the best.Using factor selection method can not only reduce the computational cost of the model,but also effectively improve the performance of the model.(4)The SHAP algorithm is used to establish the interpretable model of landslide susceptibility mapping in the study area.The global and local levels of the mapping model are explained by the summary plots,the dependence plots and the waterfall plots.The analysis results of the summary plots show that the three topographic factors of profile curvature,elevation and hydrodynamic index have the greatest impact on the landslide disaster in the study area.The waterfall plots were used to analyze the susceptibility of Shuangtu Laochang town landslide and Tielu landslide,and the analysis results were compared with the field investigation report.The analysis results show that the hydrodynamic index,rainfall condition and elevation are the main influencing factors of the two landslides,and the analysis results are basically consistent with the field investigation report.The interpretable model provides a rich supplement for the results of susceptibility mapping,which is more persuasive.