Estimation and analysis of CO2 column concentrations (XCO2) in the Yangtze River Delta of China based on multi-source data and machine learning

Carbon dioxide (CO2) is one of the most significant greenhouse gases in the atmosphere and plays a crucial role in global warming. Currently, the temporal resolution for XCO2 from the satellite is low, and the ground-based XCO2 observation station is limited. There is an urgent need for a XCO2 dataset with high temporal and spatial resolution. Consequently, based on the random forest algorithm, we have developed an optimized model for predicting XCO2 with a spatial resolution of 0.25 degrees x 0.25 degrees and a temporal resolution of 1 h for the Yangtze River Delta in 2020. The multi-source data, such as the ground-observation XCO2 from the TCCON, as well as meteorological parameters, aerosols, surface vegetation index, and emission source factors from the ERA5, MERRA-2, MODIS, and MEIC datasets, were used in this study. The results indicate that the random forest model is well-suited for predicting XCO2. Specifically, the model performs more optimally when utilizing 20 variables, including solar zenith angle, normalized vegetation index, and carbon emission data as input parameters with the prediction RMSE and R2 of 1.031 x 10-6 and 0.940. The MAE for predicted XCO2 at Xianghe and Hefei stations are 0.628 x 10-6 and 0.550 x 10-6, respectively, marking a substantial increase in accuracy compared to GOSAT data. In 2020, daily variations of XCO2 follow a pattern of higher concentrations at night and lower concentrations during the day, negatively correlating with changes in the atmospheric boundary layer height. The intermonthly and seasonal variations reveal smaller concentrations in summer and higher concentrations in winter. The minimum concentration occurs in July at 409.64 x 10-6, while the maximum concentration occurs in November at 413.11 x 10-6. Spatially, XCO2 is higher in the northern areas and lower in the southern regions, showing a negative correlation with the NDVI and a positive correlation with anthropogenic carbon emissions. The XCO2 dataset calculated in this study with continuous spatial and temporal resolutions could address the limitations of satellite products with low temporal resolution and a limited number of ground observation stations.