|The authors have made substantial revisions to the original manuscript and the new version is greatly improved. The added discussion of the variability of the averaging kernels due to clouds should be particularly helpful to readers. However, there are remaining problems with respect to the interpretation of the validation results.|
Previous satellite products for CO (e.g., MOPITT, TES, and IASI) have been validated explicitly using the retrieval averaging kernels to account for the influence of smoothing error and a priori information. This approach typically requires the availability of in-situ vertical profiles, but allows smoothing error to be separated from other sources of retrieval error. This is critical since, for example, comparisons of retrieved CO profiles with model output are not affected by smoothing error (assuming that the retrieval averaging kernels are properly used).
The validation strategy employed by Borsdorff et al is to exclusively use ground-based TCCON retrievals of CO dry-air mole fraction (at multiple sites) and to neglect retrieval differences arising from the different averaging kernels for TCCON and TROPOMI. This strategy itself is not unreasonable: A good model for this approach is reported in Cogan et al., "Atmospheric carbon dioxide retrieved from the Greenhouse gases Observing SATellite (GOSAT): Comparison with ground-based TCCON observations and GEOS-Chem model calculations," JGR, 117, doi:10.1029/2012JD018087, 2012. In that paper, the authors quantify expected differences in GOSAT/TCCON CO2 retrievals due to averaging kernel differences using the GEOS-Chem model to simulate a realistic range of CO2 profiles.
In the revised manuscript, Borsdorff et al choose to interpret the TCCON CO total column values as 'truth' and refer to a technical report by Wunch et al as the source of a claim that TCCON CO total column values are accurate to better than 4%. However, since smoothing error (due to variability of the true CO profile shape) was not specifically studied as a source of retrieval error in that report, the actual uncertainty of the TCCON CO total column values is likely greater than 4%.
To clarify these issues and emphasize their importance, the authors should make the following additional revisions.
1. Following the sentence "However, the TROPOMI CO dataset provides total column averaging kernels for each retrieval and we recommend to use them when ever possible" in Section 2, please provide clear instructions to potential users (including the relevant equation) regarding the actual method for applying the total column averaging kernel to CO in-situ datasets or model simulations.
2. The similarity of clear- and cloudy-scene validation statistics at the end of Section 3.1 might imply to readers that these two subsets are equivalent with respect to data quality and can be used interchangeably. However, the similarity in validation statistics could simply result from the selection of scenes where the under-cloud CO profile was generally consistent with the above-cloud CO profile (in terms of profile shape). This would not be the case in strong CO source regions, for example. Please re-emphasize to the readers at the end of Section 3.1 that clear-sky retrievals are always preferable (compared to cloudy-sky retrievals) because of the more consistent sensitivity to CO over the entire profile.
3. Since the validation results reported in this paper largely hinge on the absolute accuracy of the TCCON retrievals, please provide stronger quantitative evidence for the effects of smoothing error on TCCON CO retrievals, preferably from the peer-reviewed literature. If the effects of smoothing error as a source of retrieval error for the TCCON retrievals have not been explicitly studied, that should be stated in the manuscript.