Abstract: The deformation evolution process of loess landslide is often nonlinear due to many factors. A theoretical approach named WA-BT-ELM, which is based on wavelet analysis (WA), boosting regression tree (BT) and extreme learning machine (ELM), is proposed to predict loess landslide displacements. By analysis of nonlinear loess landslide time-dependent displacement dataset, the accumulation displacement data signal is decomposed into a series of sub-wavelets. Then, the importance of all the sub-wavelets to the displacement data series is computed by BT algorithm. The highly important sub-wavelets are selected to make further predictions. Furthermore, the predictive results of sub-wavelet and the original landslide displacement series are obtained through ELM algorithms. A case study of Heifangtai landslide in Gansu Province is presented to verify the predictive results. In comparison, the predictive results by using WA-BT-ELM model is faster and more accurate than those by ANN,BPNN,SVM,ELM and WA-ELM model, indicating that the WA-BT-ELM model is effective in loess landslide displacement prediction cases.

Key words: loess landslide, WA-BT-ELM model, displacement prediction, wavelet analysis, boosting regression tree, extreme learning machine