Graph attention network-based multi-source data fusion for atmospheric CO ₂ concentrations estimation

Atmospheric CO2 concentration is crucial for understanding carbon emissions, but satellite observations often suffer from extensive pixel loss. Existing interpolation and machine learning methods have been used to estimate missing CO2 concentrations, but these approaches result in coarse resolutions and fail to account for spatial correlations. To address these challenges, this study proposes a novel graph attention-based multi-source data fusion method to achieve high-resolution and high-accuracy CO2 concentration estimation. Each carbon observation is treated as a node in a graph, the method employs clustering-based subgraph partitioning and constructs an adjacency matrix based on spatiotemporal distances to effectively capture spatiotemporal information from sparse data. In the model, spatial attention is used to capture the spatial proximity of CO2 concentrations, while feature attention captures the complex correlations between variables. Furthermore, an advanced fusion module with linear layers models nonlinear relationships between multi-source features and CO2 concentrations. Experiments conducted in urban areas demonstrate the effectiveness of the proposed method (overall R2 = 0.904, RMSE = 1.109 ppm), with smoother mapping results and better identification of CO2 gradients, aiding in the localization of carbon sources. Variable importance analysis indicates that climate and pollutant concentrations contribute the most to CO2 estimation.