Among various carbon dioxide (CO₂) monitoring approaches, satellite observation and ground-based observations are two widely adopted and reliable methods. However, the former is constrained by relatively low accuracy, while the latter suffers from limited spatial coverage. Therefore, this study develops a hybrid spatiotemporal modeling method integrating two transformer networks based on heterogeneous graphs to enhance the Copernicus atmosphere monitoring service global greenhouse gas reanalysis version 4 (CAMS-EGG4) dataset by fusing orbiting carbon observatory-2 (OCO-2) satellite observations. Dynamic heterogeneous spatiotemporal graphs are constructed by integrating spatial proximity with corresponding temporal nodes, enabling an explicit representation of the intrinsic spatiotemporal dependencies in atmospheric CO₂ observations. This approach maintains the advantage of broad spatial coverage while achieving accuracy comparable to ground-based observations. The proposed model demonstrates high predictive performance, with a coefficient of determination (R²) of 0.97, a root mean square error (RMSE) of 0.99, and a mean absolute error (MAE) of 0.79. We further benchmarked the proposed model against four machine learning approaches, including random forest (RF), extreme gradient boosting (XGBoost), convolutional neural network (CNN) and graph convolutional network (GCN). Experimental results show that the dual–transformer model consistently outperforms the four approaches in terms of R², RMSE, and MAE. In addition, validation against ground-based measurements from the total carbon column observing network (TCCON) confirms that the predicted XCO₂ data exhibit substantial improvements over the original dataset. This methodological advance offers a scalable foundation for constructing refined CO2 reanalysis products, with broad applicability in atmospheric science, earth system modeling, and evidence-based climate policy development.