Daily High-Resolution CO₂ Mapping over European Urban Areas from Targeted Satellite Observations Using Machine Learning

Urban regions are major contributors to anthropogenic CO₂ emissions, yet satellite-based observations of column-averaged CO₂ remain spatially and temporally sparse, limiting high-resolution urban monitoring. This study presents a machine-learning framework for predicting daily CO₂ over European urban areas at a spatial resolution of 0.02°, integrating satellite NO₂ observations, reanalysis meteorological variables, and surface data. High-density OCO-2 target-mode observations are used as ground truth, enabling robust learning of the relationships between CO₂ and its atmospheric and surface drivers.

Two predictive scenarios are evaluated. The first, a sample-based prediction designed primarily for spatial gap filling, achieves an R² of 0.98 ± 0.00 and an RMSE of 0.60 ± 0.02 ppm using 10-fold cross-validation. The second scenario assesses spatiotemporal generalization, yielding an average R² of 0.91 ± 0.02 and RMSE of 1.17 ± 0.14 ppm for temporal transfer, and R² of 0.85 ± 0.09 with RMSE of 1.25 ± 0.20 ppm for spatial transfer across European regions. Independent validation against ground-based CO₂ measurements from the Munich Urban Carbon Column network (MUCCnet) shows strong agreement, with R² values between 0.95 and 0.97 and RMSE ranging from 0.50 to 0.72 ppm.

The results demonstrate the potential of the proposed framework to fill observational gaps and generate reliable, high-resolution CO₂ fields over urban environments and their surroundings, supporting improved monitoring of anthropogenic CO₂ emissions where accurate information is most critical.