Prediction Of Rainfall-induced Landslide Deformation Based On Deep Learning

Landslide disaster seriously threatens the safety of human life and property.Due to the complex geological conditions and large undulating terrain,the southwest area of China is prone to occur landslides.Statistics show that,except for a small part of landslides caused by earthquakes or artificial factors,80% of landslides are rainfallinduced landslides.With the arrival of the rainy season every year,with the increase of rainfall and the length of rainfall days,the rainfall-induced landslides enter a period of high incidence.In order to feasibly and effectively carry out prevention and early warning of rainfall-induced landslides,a large number of landslide monitoring equipment are arranged at hidden landslide points.The rapid development of monitoring equipment and data collection,transmission,and storage technology has promoted the research of indepth mining of landslide information.The deep neural network model in deep learning technology has become the first choice for processing complex landslide data.However,due to the complexity,variability,and randomness of the landslide,landslide prediction and forecasting still face many problems.This paper focuses on the deformation prediction of rainfall-induced landslides.First,explore the applicability of commonly used outlier elimination methods and missing value filling methods in the preprocessing of landslide monitoring data.Then,analyze the displacement data of the landslide monitoring point and perform feature engineering processing including operations such as sample set combination and sample feature selection on it.Finally,in view of the defect that deep neural networks(DNN,Deep Neural Networks)are easy to fall into the local optimum and lead to reduced accuracy,the Ada Boost algorithm in Boosting is used to improve the DNN deep neural network,and the accuracy of the prediction results is evaluated.Through the analysis and research of the above-mentioned problems,this paper has mainly obtained the following research results:(1)Combining the consistent characteristics of landslide data,first use the box chart to eliminate outliers,and then secondly filter the "outliers" eliminated from the box chart,and divide them into true outliers and the False outliers of real law of landslides.After removing the outliers,this paper discusses the applicability of several missing value filling methods in monitoring data of rainfall-induced landslides.(2)The feature engineering method of data mining is introduced to analyze the influence factors of rainfall-induced landslide.Firstly,the original rainfall data were transformed into the current day's rainfall grade,previous effective rainfall,historical average rainfall,and historical rainfall days,so that the rainfall characteristics are more in line with the mechanism of rainfall on the displacement of the landslide;Then,considering the continuity of the landslide's own state,the scatter diagram analysis of the historical slip value and the predicted value shows that the linear relationship between them is strong.Finally,the influence of the time feature on the daily displacement of the landslide is analyzed,and the month is added as a feature to the sample set.After analyzing the above factors,the feature interaction method is used to enrich the feature representation,and the tree algorithm and the Spearman correlation coefficient method are used to remove irrelevant features and redundant features.(3)Established a landslide daily displacement deep neural network prediction model,elaborated the method of determining the hyperparameters in the prediction model,and aimed at the defect that the deep neural network is easy to fall into the local optimum and resulting in low accuracy,Ada Boost algorithm in Boosting was proposed to deal with it.The results show that the prediction accuracy of the deep neural network model processed by the Ada Boost algorithm is significantly improved.In addition,by drawing a learning curve and comparing it with traditional methods,the experimental results show that the Ada Boost-DNN algorithm is more sensitive to the amount of data.As the amount of data used for training increases,the prediction accuracy gradually improves.In the case of ensuring the amount of data,the prediction accuracy is better than traditional algorithms such as decision trees,SVM,GBDT,and random forests.