Dear authors,
After a 2nd review two of the reviewers still have a number of concerns that should be addressed in the final version. Below I have posted the comments from reviewer #3 and attached are the comments made by reviewer #1.

Best regards,
Christoph Gerbig
Comments reviewer #3:
Review #2 of Lan et al., 2017.

This is the first time I am seeing a response to my original review comments, so I will respond to the author comments here. My main concern is that two of the three original reviewers requested comparisons with TCCON, and the authors have done so, but have not included this in the revised manuscript nor in the supplementary materials.

Author comments are in quotes and my responses are interspersed.

"We agree with the reviewer that it will be interesting to compare TCCON with aircraft + CT based total column CO2. The following figure can give us some ideas about this comparison. TCCON values in this figure are daily averages computed from mid-day data (10 to 15 LST). Outliers in TCCON data are filtered out, by using 3.S.D threshold of residuals (with iteration) from a quadratic fit on data for each day. However, it is not an apple-to-apple comparison between our aircraft-based column CO2 and TCCON. The sampling time and location are different. Aircraft sampling typically takes ~ 30 min to get the whole profile through downward spiral, mostly within 0.1⁰ of the site location."

I agree that what you have plotted is not an apples-to-apples comparison. I would suggest that instead, you average the TCCON data during the 30 minute aircraft sampling times instead of simply taking mid-day averages. I also do not see the need for your additional 3-sigma filtering.

"In addition, the aircraft and tower measurements are calibrated on WMO scale; however, direct calibrations of remote sensing instruments are not available. It makes more sense to compare TCCON against calibrated measurements for accuracy instead of the other way around. Our study focuses on the long-term mean spatial patterns of column CO2 over North America and the influence of large-scale transport..."

The overall scaling of the two data products is not of prime interest, here. As you mention, the long-term mean spatial pattern are primarily important, which I believe you can assess using a comparison with TCCON data. This might give you a way of assessing the CT stratosphere, and improving your product!

"For TCCON that sampling the whole column instantly at an angle, the actual sampling location, especially at high altitudes, may be far away from our aircraft inlets even though the distance on the ground is much smaller."

Near noon, this should be a less significant problem. At the very least, this is a quantifiable problem, and you have not demonstrated the problem in a quantitative manner. Also, you base much of your paper on the fact that the stratosphere does not change significantly in time or space, so this argument is not strong. (For example, you later state: “There is little spatial variation of atmospheric CO2 in the stratosphere over mid-latitude region (Andrews et al., 2001).” and “We cannot properly assign an exact number as the uncertainty for the top 1/3 of the atmosphere without routine in-situ measurements. However, spatial gradient and atmospheric variability is very small in this part of the atmosphere, and thus it is not important in our study.”)

My other requests were mostly satisfactorily dealt with.

The response to the reviewers comments are acceptable, with one exception:
Reviewer #3 comments regarding a potential sampling bias in the comparison with the BESD multi-year June-August averaged columns from Reuter et al., 2014: "SCIAMACHY only makes measurements when it flies overhead, and when it is sufficiently clear".
In your reply you show a figure ("Figure for reviewing process") column CO2 from sampling CT2005 with SCIAMACHY locations and dates, and you argue that the sampling bias is unlikely the reason for large spatial differences.
I see the following issues:
1) the sampling of CT2015 CO2 fields does not take into account any clear sky criterion (only then can columns be retrieved from the satellite) except for matching the date and location. I do see still potential for a larger sampling bias. In the model the cloud cover could well be different. Also if the sampling was not made during the same hour of the day (rather than just during the same date), a bias can arise. This should be clarified and included in section 3.5 of the manuscript.
2) The patterns in the CT2005 columns when sampling at SCIAMACHY locations/dates shown in the "Figure for reviewing process" are also not smooth, in fact in terms of smoothness they are not so different from those shown in Fig. 2a in Reuter et al., 2015, and they are more different from the smooth patterns in Fig. 11. This indicates a rather strong impact of a sampling bias, even when no additional criterion for cloudiness or hour of day is used.

Later on you state in your reply: "The reviewer seems to be confirming that, in fact, SCIAMACHY's sampling biases prevent scientists from being able to infer reliable sources/sinks." This is completely in contrast to how I interpret this reviewers comment. All the reviewer is saying is that so far you haven't been able to prove that the satellite retrievals are unphysical, as sampling biases can potentially explain this. I share this opinion, and I would think unless a sampling bias can be excluded as a reason, the language in section 3.5 needs to be toned down significantly.

Dear Authors,

The last version of the manuscript has clarified a few issues with respect to sampling bias and averaging period. However, I feel that the language used specifically in section 3.5 is not appropriate given the evidence presented. The reader gets the impression that simply based on the disagreement with CT2015 derived columns the columns retrieved from satellite observations become "unrealistic" or "unphysical", leading to a statement such as “they should not be used to derive any estimates of a European carbon sink”.

The evidence is that CT2015 uses ground-based observations with good coverage over North America, but with limited coverage over Europe, to estimate posterior fluxes. Forward modeling of CO2 based on these posterior fluxes compares favorable with aircraft profile observations over North America, but not over Europe where such profile data are not available (as also pointed out by the reviewers). Any sampling bias resulting from the sparse surface network coverage over Europe was not taken into account; such a bias was shown by Reuter et al. (2014) to lead to significantly smaller net uptake estimates over Europe. Any potential impact of structural errors in the biosphere flux model used in CT2015 on flux patterns over Europe and thus on gradients are not mentioned. When sensing sampling bias of remote sensing observations is taken into account by sampling the model at the correct locations and times, the pattern of CT2015 columns change, but this does not explain the difference to patterns in SCIAMACHY derived columns over Europe. Actual filtering using only clear sky periods based on cloud cover information from the meteorological fields used to drive CT2015 was not investigated nor was any potential effect on the CT2015 cerived patterns in CO2 columns mentioned.

The strong wording in section 3.5, specifically lines 434 – 436, is not justified given the evidence. If this is not modified to better reflect the limited evidence, I have no choice but to reject the paper.

Dear Authors,

Thank you for revising the wording in section 3.5. Regarding my second comment on cloud cover, there seems to be a misunderstanding. Also in CT2015 there is cloud cover likely affecting radiation and thus photosynthesis. Therefore a filter that would exclude cloud contamination in CT2015 fields before comparing to cloud-free scenes from SCIAMACHY BESD could be used.

Nevertheless I suggest publishing at this state.