A Deep Learning Approach Using Gated Recurrent Unit for Prediction of Landslide Displacement Based on Spatial-Temporal Features of Multi-Monitoring Points

Abstract

The existing prediction models concentrated on the displacement prediction based on one monitoring point with obvious displacement characteristics, and there have been few results considering the multi-monitoring points with unobvious characteristics. This paper proposes an improved multi-monitoring-points-based method to predict landslide displacement using gated recurrent unit (GRU) neural networks. Firstly, the weighted undirected graph and the Gaussian function are employed to propose a star topology location tensor (STPT) for extracting spatial features between the predicted points and the surrounding adjacent monitoring points. Meanwhile, the GRU is utilized to extract temporal features of monitoring data. Then the future displacement is predicted using the spatial-temporal features. By using a displacement dataset of the Baishuihe landslide, the effectiveness of the proposed method is demonstrated in the comparison with the existing models.