Uncertainties Of Landslide Susceptibility Modeling Considering Different Resolutions And Proportions Of Training And Testing Dataset And Detailed Hazard Warning

China is one of the countries with the most serious geological disasters in the world.The complex and diverse geological structural environment,the spatial and temporal distribution of different climatic and the frequent human engineering activities are the main causes of geological disasters.Landslide susceptibility map can provide specific locations from the perspective of spatial probability and rainfall threshold model can provide specific induction time from perspective of time probability.Therefore,it can realize the temporal and spatial warning of regional landslide,namely landslide hazard warning,combined landslide susceptibility with rainfall threshold model,which can provide important guidance for regional rainfallinduced landslides warning and prevention.However,in the process of landslide susceptibility modeling,the selection of different resolutions of evaluation units and different proportions of training and testing dataset will bring great uncertainty to the prediction results of modeling.In addition,the traditional critical rainfall threshold can only be qualitatively divided into rainfall threshold levels,which has poor spatial identification.Therefore,in view of the above the existing problems of susceptibility and critical rainfall threshold in landslide hazard,this study takes Ningdu County of Jiangxi Province in China as the research object to carry out regional landslide hazard warning.The main research contents and results are as follows:(1)446 landslides inventory information from 1970 to 2003 in Ningdu County are obtained.In addition,the 13 of typical environmental factors including topography,basic geology,hydrological environment and land cover are chosen based on the natural geographical characteristics and references of similar research area in this county.Morever,the distribution characteristics of landslide in interval of environmental factors and the correlation and collinearity between environmental factors are analyzed.(2)25 kinds of training datasets and testing datasets are obtained for the training and testing of random forest(RF)and support vector machine(SVM)models under the combination of two uncertain factors of different resolutions and different proportions of training and testing dataset.The results show that the RF model with the combination of resolution with 15 m and proportions of training and testing dataset with 9:1 has the highest prediction accuracy,with AUC of 0.915.Furthermore,with the decrease of resolution and proportion of training and testing dataset,the prediction accuracy of two models decreases,the mean value of susceptibility indexes increases,its standard deviation decreases,and the uncertainty increases.(3)A total of 176 rainfall-induced landslides are selected based on the occurrence dates and induced factors in landslide cataloge,and the rainfall dates of landslide are obtained based on 8 rainfall stations within and nearby study area and spline interpolation method.In addition,the temporal and spatial distribution characteristics of rainfall and landslide and the relationship between landslide and daily rainfall,early effective rainfall,continuous rainfall days and highway density are analyzed.(4)The detailed critical rainfall threshold is obtained based on the traditional critical rainfall threshold and logistic regression equation.In addition,the highway density is introduced,and and the detailed static and dynamic rainfall threshold is obtained by taking the frequency ratio of highway density as the static inducer and rainfall as the dynamic inducer.The results show that there are 18 landslides felling in the region of 3-level special attention and above in the traditional rainfall threshold and the traditional static and dynamic rainfall threshold(the traditional critical and static and dynamic rainfall thresholds),and there are respectively 16 and 12 landslides felling in the region of 70% and above in the detailed critical rainfall threshold and the detailed static and dynamic rainfall threshold(the detailed critical and static and dynamic rainfall thresholds).The detailed critical rainfall threshold has better effect compared with the detailed static and dynamic rainfall threshold.(5)The susceptibility and rainfall threshold are combined to carry out the detailed critical hazard and the detailed static and dynamic hazard(the detailed critical and static and dynamic hazard)warning.The results show that the two landslides have good warning effect in both the traditional and the detailed hazards.Two landslides fall in the area with very high susceptibility(above 0.8).In addition,one landslide falls in the area with five-level special warning(90%?100%)in rainfall threshold,and one landslide falls in the area with four-level warning(70%?90%)in rainfall threshold.Moreover,two landslides fall in the area with very high hazard(above 0.7).(6)The influences of soil erosion factor as static inducing factor on landslide are preliminarily explored.The results show that soil erosion is the most important factor in models,which can improve prediction performance of models.In addition,with the increase of soil erosion level,the landslides probability will also increase.