Multigraph Convolutional Networks for Rainfall Estimation in Complex Terrain

Accurate rainfall estimation over complex terrain is critical for science and applications concerning life and economy, but it is challenging due to the multifactorial relationship between topography, environmental parameters, and rainfall intensity. In this work, a graph convolutional neural (GCN) network-based approach named multiGCN network (M-GCN) is used to interpolate precipitation at a 30-min temporal scale. Furthermore, to enable the model to adapt to the variabilities of spatial correlation, we cluster the ground radar nodes based on their geographical information and expand the network with the multigraph mechanism. Thus, we can avoid overfitting caused by varying conditions over a wide area, and the estimation accuracy can be improved. The method was tested on ground radar-gauge precipitation data over three months on the West Coast of the United States, in 2015. The experimental result confirms that our proposed method outperforms the state-of-the-art interpolation methods. Besides interpolation capacity, the M-GCN also has the advantage of computational efficiency. The distributed graphs in the M-GCN architecture make it possible to train the networks on edge servers and the cloud.