Geospatial-temporal Heterogeneity Embedded Graph Neural Network for AERONET AOD Forecasting

Accurate prediction of Aerosol Optical Depth (AOD) is crucial for characterizing atmospheric turbidity and understanding the climate impact of atmospheric aerosols. In this study, more reliable and precise AOD data from various Aerosol RObotic NETwork (AERONET) sites are selected as the experimental data. However, improving the accuracy of predicting unstructured AOD site data is challenging without accurately quantifying the spatial and temporal relationships between sites, given the presence of geospatial-temporal heterogeneity in geographic environments. To tackle this challenge, we propose a geospatial-temporal heterogeneity embedded graph neural network (GSTH-GNN) for the simultaneous forecasting of multiple AOD site data. Furthermore, a novel geographically and temporally weighted regression based on a graph attention network (GTWRGAT) is also proposed to estimate geospatial-temporal heterogeneity on the spatio-temporal graph in this paper. Specifically, we quantify the spatio-temporal non-stationary of each site in the graph to construct a heterogeneous adjacency matrix, directing the graph convolutional network (GCN) to propagate information, while the temporal features of the sequences are captured using gated recurrent unit (GRU). The model employs inductive learning to implicitly calculate the spatio-temporal dependencies between sites, enabling the spatial and temporal learning module to capture more precise features and enhance prediction accuracy. Having evaluated 67 AERONET sites within the USA, our proposed GSTH-GNN model shows better prediction results when compared with five baseline models. Finally, we perform an analysis of geospatial-temporal heterogeneity and conduct an ablation experiment. The results indicate that leveraging a priori knowledge to quantify the spatio-temporal non-stationarity between sites for measuring their spatio-temporal dependence contributes to enhancing model performance.