Multi-task learning for landslide displacement prediction

The accurate prediction of landslide displacement has very important practical significance. At present, most studies only make individual predictions for each displacement monitoring point, ignoring the contribution of the correlation among different monitoring points to the prediction results. In this paper, a novel method for landslide displacement prediction using multi-task learning is proposed. In this methodology, the displacements of different monitoring points on the same landslide surface are simultaneously predicted due to the similarity among them. Experimental results on four landslides near Huangdeng Hydropower Station in China verify the effectiveness of the multi-task method for landslide displacement prediction. In addition, considering the characteristics of landslide displacement, a hybrid prediction method combining double exponential smoothing and multi-task method is proposed. The experimental results on four real-world landslide datasets demonstrate the effectiveness of the hybrid prediction method.