Accurate monitoring of global carbon dioxide (CO?) column concentrations (XCO2) is crucial for understanding carbon cycles and supporting climate mitigation policies. However, current methods, including satellite observations and atmospheric transport models, each face significant limitations. Satellite-based XCO2 products are hindered by limited spatial-temporal coverage and retrieval uncertainties caused by cloud interference and surface reflectance variability. Meanwhile, chemical transport models, such as GEOS-Chem, often exhibit systematic biases due to uncertainties in emission inventories and parameterizations. To overcome these challenges, we aim to develop a high-precision, spatiotemporally continuous global XCO2 dataset by assimilating multi-source satellite observations into the GEOS-Chem model using an ensemble Kalman filter (EnKF). This approach is designed to meet the urgent need for reliable, high-resolution CO? monitoring systems that can support carbon flux inversion and global carbon budget assessments.

We integrate three satellite-based XCO2 products (TanSat, OCO-2, and GOSAT) into the GEOS-Chem v14.2.3 chemical transport model using an ensemble Kalman filter with covariance localization. The assimilation system is designed to generate a global XCO2 dataset with a 3-hourly temporal resolution and a 2.0°×2.5° spatial resolution for the period from March 1, 2017 to February 28, 2018. Satellite data are preprocessed with quality screening and weighted averaging based on normalized prior uncertainties [Eqs. (3)?(6)]. Model output from GEOS-Chem is vertically integrated to obtain column-averaged concentrations [Eqs. (1)?(2)], and a 20-member ensemble is constructed using perturbed initial states to represent model uncertainty. Covariance localization is applied using a Schur product approach [Eqs. (10)?(11)] to mitigate spurious correlations in the high-dimensional state space. The Kalman gain and state update equations [Eqs. (12)?(14)] ensure physical consistency during the assimilation. The final dataset is validated against ground-based TCCON measurements from 16 globally distributed sites.

We propose and implement a data assimilation framework tailored for multi-source satellite observations, effectively addressing the challenges of data fusion and error propagation within the ensemble Kalman filter. The results demonstrate that integrating multi-source satellite data significantly enhances the spatiotemporal coverage of global XCO2 observations, which effectively fills previous observational gaps and substantially increases the volume of assimilable data (Figs. 2 and 3). Validation shows that the GEOS-Chem model generally underestimates XCO2 concentrations, with overestimations in polar regions—consistent with previous studies. By assimilating multi-source satellite observations, these systematic biases are effectively corrected: the model’s RMSE is reduced from 1.27×10^-6 to 1.19×10^-6, and the mean bias improves from -0.42×10^-6 to -0.28×10^-6 (Fig 4). Moreover, seasonal deviations are notably mitigated (Figs. 6 and 9), and the model’s performance under extreme climatic conditions becomes more consistent with actual observations (Fig. 10).

We develop a global XCO2 reanalysis dataset by assimilating multi-source satellite observations into the GEOS-Chem model using an ensemble Kalman filter. The assimilation significantly enhances spatial-temporal data coverage, reduces systematic model biases, and improves agreement with ground-based measurements. The final dataset not only preserves realistic seasonal XCO2 dynamics but also captures extreme meteorological and geographic influences more accurately. While limitations remain due to restricted satellite data availability and the potential introduction of new observational errors, we provide a solid foundation for future carbon flux inversion studies and support enhanced climate policy implementation. Further improvements can be achieved by expanding domestic satellite participation and developing higher-resolution assimilation frameworks.