Deriving Full‐Coverage and Fine‐Scale XCO2 Across China Based on OCO‐2 Satellite Retrievals and CarbonTracker Output

Due to the coarse spatial resolution, the column‐averaged dry‐air mole fraction of CO2 (XCO2) data from the CarbonTracker may be inadequate to reflect the spatial heterogeneity of XCO2. We developed a machine learning model to fill the data gaps in the Orbiting Carbon Observatory 2 satellite retrievals across China during 2015–2018, with cross‐validation R2 = 0.95 and RMSE = 0.91 ppm. Based on the gap‐filled data set, the multiyear average XCO2 was the highest in East China (405.71 ± 3.72 ppm) and the lowest in Northwest China (403.99 ± 3.47 ppm). At the national level, the multiyear seasonal XCO2 varied from 402.54 ± 3.95 ppm in summer to 406.28 ± 3.19 ppm in spring. While the XCO2 kept increasing, the rate of increase declined from 3.23 to 2.10 ppm/year. The machine learning approach is feasible for downscaling and calibrating the CarbonTracker XCO2 data. The full‐coverage and fine‐scale XCO2 data set is expected to advance our understanding of the carbon cycles. Plain Language Summary As the most abundant greenhouse gas, atmospheric CO2 is considered one of the main contributors to climate change. Climate research requires comprehensive spatiotemporal distributions of CO2. However, the sparseness of the ground‐based measurements and lack of a fine‐scale modeled CO2 data set limits our understanding of the carbon dynamics. We developed a machine learning model to fill the data gaps in the satellite retrievals of the column‐averaged dry‐air mole fraction of CO2 (XCO2) for China during 2015–2018. The XCO2 levels were the highest in East China and the lowest in Northwest China, and were the lowest in summer and the highest in spring. While the XCO2 levels kept increasing, the rate of increase slightly declined. The full‐coverage and fine‐scale XCO2 data are valuable for policy making in carbon emission management. This work affirms the necessity of modeling the XCO2 from satellites, which can advance our understanding of the carbon cycles. Key Points The machine learning model exhibited decent performance in predicting gridded daily XCO2, with cross‐validation R2 = 0.95 The machine‐learning predicted XCO2 revealed considerably higher spatial heterogeneity than the CarbonTracker output Based on the gap‐filled XCO2 data set, the average XCO2 during 2015–2018 was the highest in East China and the lowest in Northwest China