Research On Regional Landslide Disaster Early Warning Model Based On Machine Learning In Qingchuan County,Sichuan Province

Landslide disasters occur in large groups in China,with many areas and serious hazards.Regional landslide disaster early warning is an important means of effective disaster prevention and mitigation,and early warning model is the core of successful early warning and one of the important contents of landslide disaster prevention and mitigation.However,due to the complex landslide inducing mechanism in the study area,the traditional regional geological disaster warning model has problems such as limited precision and insufficient refinement due to the lack of big data of investigation and monitoring and the lack of analysis methods.In recent years,with the rapid development of big data and artificial intelligence technology,machine learning technology has been widely used in geological disaster early warning and achieved good results.In this paper,a regional landslide disaster early warning model based on machine learning is systematically studied.Taking Qingchuan County,Sichuan Province as an example,a regional landslide disaster early warning model of Qingchuan County is constructed based on the recent ten years of geological and meteorological data,and the main research contents are as follows:(1)The development characteristics and geological environment conditions of landslide disaster in Qingchuan County are analyzed,and the spatial and temporal distribution rules are analyzed.27 indexes of 4 influencing factors,such as topographies,geological conditions,environmental conditions and human engineering activities,are selected.Eleven geological environment factors are included,including slope,slope aspect,elevation,landform type,stratigraphic lithology,historical disaster point,distance from fault,distance from river,average annual rainfall,distance from housing,distance from road,etc.Sixteen rainfall factors,such as daily rainfall and daily rainfall in the previous 1?15 days,were involved in the construction of the early warning model.(2)Based on the integration and data cleaning of the results of geological disaster investigation and precipitation monitoring in Qingchuan County,the regional landslide disaster training sample set of Qingchuan County was constructed.The sample set includes 27 input characteristic attributes such as geological environment and rainfall,and 1 output characteristic attribute,covering all samples in Qingchuan County in recent 9 years(2010 ? 2018),with a total number of 1826 samples(including 613 positive samples and 1213 negative samples).(3)Based on logistic regression model KNN nearest neighbor model decision tree model random forest model support vector machine model multi-layer perceptron neural network model six machine learning methods to build landslide disaster warning model in Qingchuan County.The sample set was trained by five-fold cross validation,and the hyperparameters of each model were optimized by Bayesian optimization algorithm.The accuracy and generalization ability of the model were verified by the ROC curve and AUC value of the accuracy confusion matrix learning curve.By comparing the accuracy of various models,the method of random forest early warning model is more suitable for the local landslide disaster early warning in Qingchuan County than other models with the highest accuracy.(4)When carrying out regional landslide actual early warning,27 feature attributes of each early warning unit in the research area are input according to the feature attribute format of training samples,and the pre-learned and trained model is called to output the occurrence probability of landslide disaster,and the early warning level of landslide disaster is determined in segments according to the output probability.When the output probability is within the interval [40%,60%),a yellow alert is issued;when the output probability is within the interval[60%,80%),an orange alert is issued;when the output probability is within the interval [80%,100%],a red alert is issued.(5)Taking the actual warning on June 26,2018 as an example,the random forest model trained in the earlier period was loaded.Based on the warning grade divided by the output probability of the calculation model and verified by the actual occurrence of landslide disasters,the results show that 100% of the disaster points are within the scope of the warning area,among which 70.6% of the landslides fall within the red warning area,17.6% within the orange warning area and 11.8% within the yellow warning area.