Training enhanced reservoir computing predictor for landslide displacement

Landslide early warning systems can be implemented based on the monitoring and prediction of landslide displacements. The internal mechanisms of landslides are very complex, and precise mechanistic models of landslides are difficult to obtain; therefore, data-driven models are usually applied. From the perspective of dynamic system theory, landslide development should be considered a dynamic process. Because traditional models, such as feed-forward neural networks, can only express static relationships, the applicability of these static models is quite limited in landslide prediction tasks. In our study, recurrent neural networks are established and trained into dynamic predictors of landslide displacement using a training algorithm named reservoir computing. A shortcoming of the reservoir computing algorithm is that the models will not perform well with insufficiently large training sets. Because practical landslide displacement recordings are usually quite short, interpolation techniques are employed to enhance the training of the dynamic predictors. The method proposed in this paper is applied to predict the developments of three landslides in the Three Gorges area, and the predicted values are close to the actual measurements. Improvements relative to sophisticated static predictors and basic dynamic predictors are obvious. (C) 2014 Elsevier B.V. All rights reserved.