| Landslides often cause enormous casualties and property losses of local community every year in our country,and the situation of landslide mitigation and prevention is severe.Due to the construction of the Three Gorges Project,the Three Gorges Reservoir Area has become the hardest hit area of landslide disaster in China.For the purpose of mitigating landslide disaster,it is essential to establish prediction model able to make landslide displacement prediction.Based on literature review,this paper proposes a framework with deep learning and statistical learning to predict reservoir landslide deformation.The framework contains six research topics,such as landslide displacement prediction,landslide displacement interval prediction,short-term deformation trend prediction,seasonal displacement identification,tail-correlation analysis for thresholds computed,and spatial-temporal prediction.The effectiveness and utility of the proposed data-driven framework have been confirmed with the landslide case studies(e.g.,Muyubao landslide,Baishuihe landslide,and Shuping landslide)in the region of the Three Gorges Reservoir.(1)Time-series analysis of landslide deformation data.First,the wavelet analysis is applied to decompose the landslide instant displacement into seasonal and residual components.Next,the seasonal characteristics of all components obtained through ACF and PACF.Last,all positive lags of triggering factors are obtained based on Pearson’s correlation coefficient.The calculations show the seasonal component and displacement increments are periodic,but the residual component is irregular.So,the displacement increments are selected as output of prediction model.(2)Landslide displacement prediction and landslide displacement interval prediction.First,a comparison model is conducted using multiple machine learning algorithms and deep learning algorithms,and the best one is chosen as “component”.Then,The Elastic Net is selected as the packaging algorithm for integrated algorithm.Next,forward-chaining nested cross-validation is performed to optimize the integrated algorithm parameters.At last,landslide displacement interval prediction is conducted with uncertainty concept for precipitation.The calculations show the EN-LSTM-RNN algorithms can accurately predict landslide displacement,and the accuracy of interval prediction is 100%.(3)Landslide short-term deformation trend prediction.In this section,short-term deformation trend prediction.Based on case studies,the predicted instant future displacement is measured by HE.The relative tendencies of displacements to converge to zero or to clustering to a significant non-zero value have been computed.(4)Seasonal displacement identification and tail-correlation analysis for thresholds computed.The EWMA control chart is developed to monitor and identify the seasonal faster displacement.And Copula models are selected to fit the predicted displacement and the major triggering factors.Based on case studies,the accuracy of identification is 100%.The Gumbel-Hougaard Copula model performs best,which indicates strong upper-tail correlation between the triggering factors and displacement values.Thresholds for the triggering factors also be obtained by monitoring the landslide moving patterns with large displacement values.(5)Spatial-temporal prediction for landslide prediction.In general,landslide equipped with global positioning system(GPS)points measuring its deformation in a fine spatially distributed scale,the kinematic of those points are connected.So,the analysis of the spatial-temporal prediction is an interesting topic.At first,the spatial-temporal correlation of landslide deformation is obtained by spatial-temporal matrix.Then,the spatial-temporal prediction model is proposed with Bayesian regression network and autoregressive(AR)models.Its advantages include a simple theoretical basis,cheap computation cost,and high general.At last,the probability distribution function of displacement increments is obtained,the thresholds of p = 0.95 is computed.The calculations show the proposed spatial-temporal prediction model can forecast landslide deformation with good accuracy. |
