China has a vast land area and its altitude is gradually decreasing from west to east.Complicated geological and climatic conditions lead to frequent geological disasters.As the most important geological disaster,landslide has been threatening the economic development and ecological environment of China.Occurrence of landslides in China,mainly small and medium-sized landslide mostly distributed along the southeast hilly region,the Yangtze river Yellow River area,and the Qinghai-Tibet plateau,because of the complex mechanism of landslide development,most of the landslide is often accompanied by debris flow and collapse and other geological disasters occur together,to roads,houses and land and other damage.Jiangxi province has a wide distribution area of mountains and hills,with extensive basins and valleys and slightly plains.The lithology of the strata is mainly magmatic rock,metamorphic rock and clastic rock.Its unique geological conditions lead to frequent landslides in this region.Landslide susceptibility prediction is one of the mainstream prevention and control methods for landslide disaster at present.The existing literature research on landslide susceptibility is mainly based on the basic data provided by the Land and Resources Bureau and external factors of field exploration and investigation,such as climate conditions,geological structure conditions,soil and vegetation distribution and human engineering activities.To conduct a series of studies to assess the potential for landslides in the area.Landslide susceptibility prediction accuracy affected by many factors,such as landslide logging information is accurate,artificial drawing landslide data whether there is any deviation,conditional factor combinations and whether there is any error,machine learning model is unreasonable parameter Settings,etc.,these factors constitute the liability to predict the uncertainty in the process of modeling,The accuracy of predicting landslide spatial probability and the reliability of landslide prevention and control measures are greatly weakened.Therefore,how to reduce the uncertainty in landslide susceptibility prediction modeling is the focus of current research.Existing landslide susceptibility prediction research there is no room for landslide boundary expression to achieve enough attention,and different landslide boundary expression will bring different results for liability prediction,even if the default of the shape of the landslide itself as polygon,doesn’t mean that using polygon as landslide boundary can obtain the optimal solution,The coupling degree of landslide boundary with geographical conditions and selected conditional factors should also be considered.In addition,the selected conditional factors are directly default to error-free and accuracy.These fuzzy data bring great uncertainty to the susceptibility prediction and greatly reduce the prediction accuracy.Therefore,based on Ruijin City of Jiangxi Province,this paper carries out prediction research on regional landslide susceptibility.Relevant methods and research results are as follows:(1)Information of 370 landslides in Ruijin city from 1970 to 2005 was obtained,and the landslide boundary expression was divided into polygons,points and buffer circles.Based on the physical geographical characteristics of the city and related literatures of similar research areas,10 conditional factors such as landform,formation lithology,hydrological environment and land cover were selected and the distribution characteristics,correlation and collinear correlation among the factors in each conditional factor interval were analyzed.(2)The frequency ratios of conditional factors under different landslide boundary expressions were obtained and combined into training and test sets respectively,which were used for the training and test of multi-layer perceptron(MLP),support vector machine(SVM)and random forest(RF)models.In addition,the accuracy prediction(ROC curve)and the susceptibility index distribution index were used to analyze the vulnerability modeling uncertainty,and finally the polygon was determined to be the most suitable landslide boundary expression method for landslide susceptibility prediction in Ruijin City.(3)in determining the polygon boundary expression in landslide is used after sex,give full consideration to the selected 10 of the conditional factors exist the possibility of error,add conditional factors,respectively,5%,10%,15% and 20% of random error,the formation lithology,the discrete conditional factor without error,and calculate the error rate under the conditional factor of frequency ratio,This process was repeated for five times to reduce the randomness and contingency.The results showed that the frequency ratios of conditional factors in the five simulations were roughly similar.In all machine learning models,the greater the random error,the lower the prediction accuracy.(4)Low-pass filtering was performed on the original conditional factor,and the factor was finally filtered twice through cross verification,with the purpose of smoothing the conditional factor and eliminating the extreme variable values in some factors.The frequency ratio of the filtered conditional factor was calculated five times and taken as the input variable of the model.The five results show that the prediction accuracy has been improved in different models,especially in the random forest model.In addition to the criterion of AUC value,the distribution law of landslide susceptibility index,the distribution map of landslide susceptibility and the relative weight of conditional factors all show that low-pass filtering can reduce the uncertainty of landslide susceptibility prediction.(5)In addition to low-pass filtering,which can reduce the error of conditional factors and improve the prediction accuracy,this paper also proposes many measures to improve the accuracy of conditional factors combined with other literature,such as optimizing the instruments for collecting data in field exploration and adding infrared light wave,lidar and remote sensing technologies to achieve more accurate data interpretation.In addition,the abundance of conditional factors such as noctilucent data,population density and human engineering activities such as highways should be increased,and the combination of factors should be appropriately changed to reduce the influence weight of a factor on susceptibility prediction,so as to further improve the accuracy of conditional factors. |