Reconstructing long-term (2003-2019) global high-resolution XCO2: bridging observational gaps with machine learning

Abstract

Atmospheric carbon dioxide (CO2), a long-lived and well-mixed greenhouse gas, is a key driver of global warming. Accurate, long-term monitoring of its spatiotemporal variability is essential for understanding carbon dynamics. While the Orbiting Carbon Observatory-2 (OCO-2) satellite provides one of the most precise column-averaged CO2 (XCO2) measurements, its limited spatial coverage and short record since 2014 constrain long-term global analysis. Many studies thus highly rely on chemical transport models (e.g. Copernicus Atmosphere Monitoring Service (CAMS) and CarbonTracker) when applying machine learning (ML) approaches. However, their coarse resolutions often lead to spatial smoothing. In this context, we present a novel ML-based framework based on residual learning with the Light Gradient Boosting Machine (LGBM) to reconstruct global, gap-free XCO2 at 0.1 degrees resolution for the period 2003-2019. By explicitly modeling the residuals between high precision OCO-2 observations and the coarse resolution CAMS-EGG4 reanalysis, the proposed framework mitigates spatial smoothing effects and enables the extension of XCO2 estimates beyond the temporal coverage of the OCO-2 mission. The resulting product was strictly validated through internal cross-validation (random, spatial, and temporal) and external in situ validation, showing strong agreement with OCO-2 satellite observations (R2 = 0.93-0.96, RMSE = 0.80-1.11 ppm) and ground-based measurements (R2 = 0.98, RMSE = 1.17 ppm), respectively. Compared to CAMS-EGG4, the LGBM-based XCO2 product also outperforms by offering higher accuracy and resolving the spatial smoothing limitations caused by its coarse resolution. By bridging gaps in satellite data across space and time, this high-resolution XCO2 product enhances applications in climate research, emission source attribution, and greenhouse gas policy assessment.