Abstract: A theoretical approach to predict landslide displacements, in which the support vector machine (SVM) method is coupled with the empirical mode decomposition (EMD) in signal processing, is suggested through the multimode SVM function regression modeling. On the basis of the historically recorded data of displacements for a slope, several intrinsic time modes for the evolutionary of displacements are obtained in the process of landslide forming by using EMD method; and they are components of statistical learning samplings with multimode information, determining the multiscale adaptive information of slope displacements varying with time. Corresponding to the information of displacement evolutionary in each empirical mode, the multimode SVM modeling method is introduced; and then the estimations of landslide displacements are obtained by the composition summing the results of slope displacements from different empirical modes. The theoretical results calculated by the proposed approach based on the monitoring data of Wolongsi new landslide and Xintan landslide show that the applications of the SVM method coupled with the EMD method, comparing with those of the genetic algorithm neural network method, have a more powerful ability for landslide displacement prediction; and the theoretical estimations are identical with the monitoring data very well.

Key words: rock and soil mechanics, landslide, displacement, support vector machine, empirical mode decomposition