Spatiotemporal prediction of landslide deformation based on temporal graph convolutional networks

Abstract

The displacement-time curves of landslides accurately reflect their movement status. Precise prediction of landslide deformation is the key to successful early warning. Machine-learning techniques have been used to predict the deformation of individual landslide monitoring points with desirable results. However, previous studies have not considered the spatial correlation between the monitoring points arranged in the horizontal and vertical profiles. Based on the deep learning model of a temporal graph convolutional network (T-GCN), a feasible solution was provided to accurately predict the overall deformation of landslides. To capture spatial and temporal correlations simultaneously, this study proposed a T-GCN spatiotemporal prediction method that considers the temporal correlation effects of the external factors inducing landslide deformation and generates comprehensive prediction results. The model combines the characteristics of a graph convolutional network (GCN) and gated recurrent unit (GRU). The GCN was used to determine the spatial correlations between landslide monitoring points, whereas the GRU was used to capture dynamic changes in displacement over time. The T-GCN model was then applied to predict the spatiotemporal deformation of the Dawuchang landslide in the Three Gorges Reservoir area. The experimental results demonstrate that the T-GCN model provides a novel solution for spatiotemporal deformation prediction of landslides and can effectively predict the overall displacement of landslides.