Spatiotemporal Interpolation Using Graph Neural Network

Abstract

Spatiotemporal interpolation is a widely used technique for estimating values at unsampled locations using the spatiotemporal dependencies in observations. Classic interpolation models face challenges, however, in dealing with the inherent nonlinearity and nonstationarity of spatiotemporal processes, particularly in sparse and irregularly sampled regions. To overcome these issues, we propose a novel model for spatiotemporal interpolation based on machine learning and graphs, called graph neural network–based spatiotemporal interpolation (GNN-STI). Our approach employs a locally stationary diffusion kernel to capture complex spatiotemporal dependencies in both sample-rich and sample-poor areas using a spatiotemporal Voronoi-adjacency graph structure. We evaluate the performance of GNN-STI against four baseline models using two experiments: a simulation experiment with a sample-rich simulated data set, and a real-world PM2.5 experiment involving both sample-rich and sample-poor areas across China. Experimental results demonstrate that GNN-STI provides accurate interpolations with high efficiency in both experiments compared to the baseline models. Therefore, our research presents an effective and practical model for spatiotemporal interpolation in various situations.