Seamless Daily XCO2 Mapping Across China From OCO-3 Observations via a Hybrid Framework of Transformer and Geo-Implicit Neural Representation

Abstract

The reconstruction of spatiotemporally continuous XCO2 is essential for the evaluation of regional carbon neutrality. However, the inherent sparsity and irregularity of carbon satellite observations restrict seamless daily mapping. Addressing the challenge of modeling complex spatiotemporal dependencies from such data, this study develops a hybrid framework coupling a transformer with a geo-implicit neural representation (Geo-INR) to reconstruct 0.05° seamless daily XCO2 across China from 2022 to 2024. Within this framework, the multikernel inducing spatiotemporal encoder (MKISTE) encodes every spatiotemporal coordinate into a latent vector, allowing the model to capture spatiotemporal dependencies effectively. The latent vector is further incorporated into the Transformer backbone to adaptively integrate orbiting Carbon Observatory-3 (OCO-3) observations with multisource auxiliary variables, including meteorological data and Earth embedding. Quantitative validation against OCO-3 observations reveals that the proposed framework exhibits excellent predictive performance on the test set ( $\mathrm{R}^{2} = 0.961$ ) and root mean square error (RMSE) = 0.723 ppm), and maintains high consistency ( $\mathrm{R}^{2}\gt 0.8$ ) with ground-based TCCON data. In addition, assessments of XCO2 mapping performance confirm that the reconstructed results capture reasonable regional patterns as expected and effectively characterize fine-scale carbon dynamics. Analysis based on the reconstructed fields shows an annual growth rate of 2.350 ppm yr−1 in China over the study period. The proposed hybrid framework and the produced seamless daily XCO2 dataset offer robust data support for refined regional carbon management.