Remote Sensing, Vol. 14, Pages 3899: Estimating Global Anthropogenic CO2 Gridded Emissions Using a Data-Driven Stacked Random Forest Regression Model (Remote Sensing)

The accurate estimation of anthropogenic carbon emissions is of great significance for understanding the global carbon cycle and guides the setting and implementation of global climate policy and CO2 emission-reduction goals. This study built a data-driven stacked random forest regression model for estimating gridded global fossil fuel CO2 emissions. The driving variables include the annual features of column-averaged CO2 dry-air mole fraction (XCO2) anomalies based on their ecofloristic zone, night-time light data from the Visible Infrared Imaging Radiometer Suite (VIIRS), terrestrial carbon fluxes, and vegetation parameters. A two-layer stacked random forest regression model was built to fit 1° gridded inventory of open-source data inventory for anthropogenic CO2 (ODIAC). Then, the model was trained using the 2014–2018 dataset to estimate emissions in 2019, which provided a higher accuracy compared with a single-layer model with an R2 of 0.766 and an RMSE of 0.359. The predicted gridded emissions are consistent with Global Carbon Grid at 1° scale with an R2 of 0.665, and the national total emissions provided a higher R2 at 0.977 with the Global Carbon Project (GCP) data, as compared to the ODIAC (R2 = 0.956) data, in European countries. This study demonstrates that data-driven random forest regression models are capable of estimating anthropogenic CO2 emissions at a grid scale.