A Study Of An Multistage Model Of Landslide Deformation Prediction Based On Trigger Factors Analysis

The Three Gorges Reservoir is one of the areas endangered severely by landslides, amid that not only the Three Gorges Dam but also the immigrant project are threatened by the landslides. The monitoring and warning affairs become the primary work since2009when the Three Gorges Reservoir had finished impoundment. The statistics show that nearly50%of the reservoir-induced slope failures take place during the first impoundment and the other mainly during the first3-5years after the dam construction. It was foreseen that this type of problem would be worrying in the Three Gorges reservoir, especially considering the annual variations of water levels resulting from flood control measures. Thus researches on methods of landslide deformation prediction can not only fill the gap in this research field, with theoretical significance, but also have an important guiding role and practical significance.Meanwhile, according to deploy of the State Council, the warning project of geological disasters have been carried on since2003in which255landslides are under professional monitoring, consequently mass data were collected include deformation data, rainfall data, ground water data as well as reservoir water level data. Hence, there is an urgent need to explore a method which is available for mass data processing.The interactions between factors that affect a landslide or a slope are complex multi-factorial, vague, random and often difficult to describe mathematically, imposing a challenge for predicting the stability a slope. Constructing models with external factors solely can overcome these obstacles, however, huge volume of data is the precondition of this approach.To sum up, constructing models to foresee landslide displacement by means of data mining technique with mass collected data is both obligatory and feasible. At present, several productive quantitative and qualitative model have been reported in the field of landslide displacement prediction, unfortunately many trials indicate that neither quantitative nor qualitative models are qualified for such a complex and difficult problem.In this paper, data mining technique was adopted to process the mass data of landslide monitoring, also several intelligent algorithms were introduced to analyze the problem with quantitative and qualitative means simultaneously. Eventually, a multistage landslide displacement prediction were implemented with proposed Environmental Response Index(ERI).The results show that CART as well as ERI obtained the optimal precision in different stage respectively. Based on this opinion, a multistage model of landslide deformation prediction based on trigger factors analysis was established. The test shows that the proposed model overpass traditional ones in prediction accuracy, and it meet requirement of practical application.Through analysis and study of above mentioned issues, the achieved results are as follow:(1) The landslide growth stage feature is acquired by using set-type data firstly, and at this stage rough set get the best result amid support vector machine, BP neural network, C5.0decision tree and logistic regression. The overall accuracy of result achieved by rough set reaches96.50%. The tests also show that rough set is an efficient tool of data reduction for landslide displacement modeling. At the same time, the Baijiabao landslide was thought to be a reservoir water induced landslide prefer to rainfall induced one as it is proved that reservoir water fluctuation play a obviously more significant role in triggering deformation than rainfall. Thus, it is advised that more attention should be paid if the reservoir is under impoundment or releasing flood.(2)The tests indicate that different algorithm should be adopted according to given landslide growth stage information. The accuracy analysis says CART is suitable for modeling of rapid deformation stage, while ERI which proposed in this paper is appropriate for other stages. The tests also shows that the comprehensive model proposed in this paper is superior to traditional methods such as time series analysis, SVM and BP neural network. And it also reaches the requirement of practical use.