Full-coverage and fine-scale spatiotemporal retrieval of XCO2 in China based on OCO-2 satellite observations

Accurate monitoring of atmospheric CO is crucial for understanding climate change, but satellite observations from missions like OCO-2 have significant spatial and temporal gaps. This study addresses this challenge by developing a high-resolution, full-coverage dataset for a key emission region. We generated a 1km monthly column-averaged CO (XCO ) dataset for China from 2015 to 2021 using a comprehensive set of predictors, including meteorological,vegetation, and anthropogenic activity-related data. We compared four machine learning models (Random Forest, AdaBoost, ANN, and CNN) and selected the best-performing AdaBoost model, validated using 10-fold cross-validation and independent ground-based TCCON data. The AdaBoost model demonstrated superior performance (CV R = 0.85), and the final product showed high accuracy against TCCON measurements (Overall R = 0.93, RMSE = 1.19 ppm), demonstrating competitive or improved accuracy compared with existing products where directly comparable. The reconstructed data reveals significant spatiotemporal heterogeneity, with high XCO concentrations in China’s major industrial and urban agglomerations, a robust seasonal cycle, and a consistent upward trend. Feature importance analysis confirmed that anthropogenic emissions, meteorological conditions, and biospheric activity are the primary drivers of XCO distribution. This study provides a reliable, high-resolution XCO dataset that is valuable for monitoring China’s carbon dynamics, verifying emission inventories, and informing climate mitigation policies.