Spatiotemporal estimation of near-surface CO₂ concentrations over the global continent based on hybrid machine learning model

Understanding the global distribution of CO₂ concentrations is crucial for studying the sources and sinks of CO₂. Therefore, it is necessary to conduct spatiotemporal estimation of CO₂ over the global continent. This study proposed a hybrid machine learning model consisting of mix attention (MA) module, deep forest (DF) module and LightGBM to estimate near-surface CO₂ concentrations. This model integrates satellite data, meteorological reanalysis data and ground station measurements to estimate near-surface CO₂ concentrations and generate a monthly CO₂ concentrations dataset with a resolution of 0.1° over the global continent. The results show that the model performs well in sample-based cross-validation, with R², MAE, and RMSE of 0.93, 1.45 ppm, and 2.40 ppm, respectively. Station validation result proves that the model has good generalization ability, with an average RMSE of 1.95 ppm. The MA module and DF module effectively enhance the model’s performance, with an improvement in R² by 0.03 and reductions in MAE and RMSE by 0.285 ppm and 0.460 ppm compared to LightGBM. Compared to XCO₂, the spatiotemporal variation of generated CO₂ data is more distinct, aiding research on carbon sources and sinks.