A global, temporally continuous, hourly CO2 concentration dataset at GAW monitoring stations for 2000–2024: Enabling fine-scale carbon dynamics analysis
Abstract. Global Atmosphere Watch (GAW) carbon dioxide (CO2) observations provide high-precision constraints on surface CO2 variability, but many station records remain temporally incomplete because of flask sampling, instrument interruptions, and quality-control exclusions. Such gaps can limit the representation of diurnal variability, bias daily, monthly, and annual mean CO2 estimates, and reduce the ability to detect short-lived events. Here, we present a global, temporally continuous hourly reconstruction dataset of GAW station CO2 concentrations for 2000–2024. The dataset was generated using a two-stage reconstruction framework that integrates GAW observations with Carbon Tracker (CT) and Copernicus Atmosphere Monitoring Service (CAMS) CO2 fields, together with anthropogenic, vegetation, meteorological, and spatiotemporal predictors. In the intra-day gap filling stage, missing hours on days with at least one valid observation were reconstructed using hour-pair models to produce complete daily profiles. In the full-sequence reconstruction stage, the remaining periods with no valid observations were reconstructed using a globally trained initial prediction model followed by station-specific post-correction. The framework showed robust validation performance at the hourly scale, with R2 = 0.96 and RMSE = 3.65 ppm for intra-day gap filling and R2 = 0.93 and RMSE = 4.76 ppm for full-sequence reconstruction. Across all stations, the mean filling rate was 68.9 %, and the mean reconstruction uncertainty was 2.80 ppm. Applying the World Data Centre for Greenhouse Gases (WDCGG) global analysis framework showed that reconstructed global monthly mean CO2 concentrations were, on average, 0.54 ± 0.21 ppm higher than those derived from the original GAW records, indicating that incomplete temporal coverage can measurably affect global monthly estimates. Compared with grid-cell CT and CAMS simulations, the reconstructed dataset retained stronger station-scale diurnal variability and captured CO2 enhancements during the 2023 Canadian wildfire period. This dataset can support long-term assessments of global and regional surface CO2 changes, fine-scale analyses of CO2 variability, and regional carbon emission inversion and evaluation. The dataset is available at https://doi.org/10.5281/zenodo.21306261 (Mi et al., 2026).
This study presents a two-stage algorithm for gap-filling of station-based CO₂ monitoring records to produce temporally continuous hourly CO₂ concentrations. The proposed methodology is novel and has the potential to be extended to other variables with missing high-frequency (hourly) observations, such as air pollutants and meteorological parameters. The resulting data product also demonstrates substantial scientific value.
The only concern I have is whether this study has conducted rigorous temporal and spatial cross-validation. Based on my understanding, the paper only applied the chronological split temporal extrapolation validation to Stage 2. Overall, there is a lack of temporal and spatial extrapolation validation, and Stage 2 also lacks spatial extrapolation validation.
The Introduction mentions that "Many stations rely on flask sampling, in which air is collected periodically and analyzed offline, typically at weekly or monthly intervals," and Figure 3 shows that many stations have filling rates close to 100%, which implies that the model requires substantial spatial extrapolation. Regarding temporal extrapolation, it is unclear whether there are many months or days during which only a few stations have available observations. This needs further clarification to determine whether month-based or day-based temporal extrapolation validation is warranted. If there are certain months when almost all stations are missing data, then in addition to temporal extrapolation validation, even more rigorous spatiotemporal extrapolation validation would be required—for example, holding out all January data from Station A for validation while training on data from all other stations for February–December. Similarly, if there are certain days when nearly all station data are missing, then station-day-based spatiotemporal extrapolation validation would be needed.
Moreover, such spatiotemporal extrapolation validation should be conducted for both Stage 1 and Stage 2.