Research On Data Mining Method And Application Of Main Control Factors Of Collapse,Landslide And Debris Flow

Collapse,landslide and debris flow are three major geological disasters in China.Amid the complex terrain,rich groundwater and rainfall as well as the increase of engineering construction in southwest China,the regional geological disasters occur frequently.As a result,people's life and property safety are constantly threatened.In this context,based on determined control factors of these disasters,the thesis established the corresponding prediction and evaluation models to provide favorable theoretical basis and data support for the prevention and control of geological disasters.In the current thesis,Wushan and Wuxi,the typical areas of collapse,landslide and debris flow in southwest China,were taken as the research areas.First of all,GIS was used to build the initial evaluation system.Moreover,evaluation system of the main control factors was constructed by multi-model big data mining methods,including geodetector method,CART decision tree and random forest.Finally,based on these evaluation systems of the main control factors,support vector machine prediction model of collapse,landslide and debris flow was established.The main research contents and achievements are as follows:(1)The evaluation system of collapse,landslide and debris flow was built,based on the primary factors.Based on the influence of topography,regional hydrology,geological structure,natural vegetation,regional meteorology and human activities,an evaluation system of29 primary factors was made.Through recursive analysis,the author got the evaluation system of main control factors with 14 disaster-inducing factors,and that of landslide comprised 19 disaster-inducing factors.For debris flow,the evaluation system of main control factors consisted of 12 disaster-inducing factors.(2)Based on the multi-model,the factor importance of collapse,landslide and debris flow was analyzed.The importance ranking of factors from three machine learning algorithms:geodetector CART decision tree and random forest,was compared and analyzed.Based on the factor importance analysis of geodetector,the first main control factors were groundwater type,elevation and stratum age;the second main control factors were POI core density,stratum age and river network density;the third main control factors were multi-year average rainfall,groundwater type and comprehensive slope.(3)The interaction and correlation of factors were investigated.There was a strong negative correlation between the section curvature of the collapse and the slope position,slope shape as well as micro landform,and then a strong positive correlation between the plane curvature of the collapse and the slope position,slope shape and micro landform was found.Moreover,a strong negative correlation existed between the plane curvature and the section curvature of the collapse.There was an extremely negative correlation between the three factors of the landslide: the sediment transport index,the hydrodynamic index as well as the section curvature,and the plane curvature,the slope shape,the slope position as well as the micro landform,respectively.The correlations among debris flow factors were same to those of landslides and collapses.Furthermore,in landslide disasters,the slope shape had antagonistic effect on 18 factors,the topographic relief had antagonistic effect on 15 factors,and the slope position had antagonistic effect on 13 factors.(4)The prediction models of collapse,landslide and debris flow were established.According to 29 primary factors,the CART decision tree prediction model and the random forest prediction model of these disasters were established.Next,the SVM prediction model was made based on the main control factors.Ultimately,through the SVM prediction model of collapse,landslide and debris flow,the rationality of the selection for main control factor was verified.(5)Models were evaluated by confusion matrix and ROC curveThrough the confusion matrix and ROC curve,this thesis analyzed CART decision tree prediction model,random forest prediction model and SVM prediction model based on main control factors.As the results showed that both the prediction effect and stability of SVM prediction model were better than CART decision tree prediction model and random forest prediction model.Specifically,the prediction model of SVM was all better in landslide,collapse and debris flow.The AUC values of all samples were 0.925,0.912 and 0.992 respectively,and meanwhile the accuracy was 0.954,0.945 and 0.985 respectively,and the recall rates were 0.889,0.871 and 1.000 respectively.