ST ‐YangNet: Spatiotemporal Yang Chizhong Graph Convolutional Network for Nonstationary and Nonlinear Spatiotemporal Interpolation

ABSTRACT

Spatiotemporal interpolation is a fundamental technique employed to mitigate data scarcity when analyzing spatiotemporal patterns of geographical phenomena. However, modeling nonstationary and nonlinear variations in spatiotemporal processes remains a critical challenge, particularly under conditions of sparse sampling. To address this challenge, this study introduces the multi‐scale compound variable theory, which posits that a geographical variable can be decomposed into global regular and local irregular components. The global regular component represents a nonstationary and smooth trend, while the local irregular component represents local nonlinear structures (e.g., hotspots and outliers). Guided by this theory, we develop a novel model for nonstationary and nonlinear spatiotemporal interpolation by coupling the spatiotemporal Yang Chizhong method with an inductive graph convolutional network (ST‐YangNet). Leveraging the distinct properties of these components, ST‐YangNet employs the spatiotemporal Yang Chizhong method to construct the global regular component and builds an inductive graph convolutional network to learn the mapping relationship between the constructed global regular component and the corresponding observed values for modeling the local irregular component. Experiments on both simulated and real‐world datasets demonstrate that ST‐YangNet outperforms five baseline methods in terms of interpolation accuracy and stability, achieving significant improvement in spatiotemporal interpolation with sparse sampling.