General comments

The study of Payne et al. (2021) presents new PAN retrievals from CrIS obtained by optimal estimation, which fits perfectly in the scope of AMT. The retrieval methodology is clearly explained and the validation against aircraft measurements is carefully led. This study demonstrates the ability of CrIS to measure PAN and capture its variability even in background conditions.

Therefore, I recommend the publication of this paper in AMT. However, I have a few comments and questions listed below.

Specific comments:

- Abstract and Section 5 (“Discussion and Conclusion”): “This bias does not appear to show a dependence on latitude or season.”

This statement does not look so clear to me, looking at Fig.6. Without any dependence, we would expect a randomly distributed differences (around the mean bias of -0.08 ppbv), while it looks like in Atom1 and 3, the biases are larger for latitudes 5-40°N, and in Atom2 and 4, larger for about latitudes 30°S-10°N (although less clear here, due to smaller sampling). This could have been due to a systematic bias effect and not a latitudinal effect if CrIS was showing a systematic bias with the aircraft, but the comparisons show a slope of 0.99, so only a constant absolute bias is expected (and furthermore Fig.3 is not showing maximum of PAN at these latitudes, except maybe for Atom2). Can the authors explain more why they assume that there is no dependence on latitude? What could be the reason of the larger bias in Atom1/3 in 5-40°N?

- Section 3.1 Retrieval algorithm and strategy

You mentioned in Sect. 4, that you also used Rodgers 2000 for deriving estimated uncertainties (noise, model parameters, ...), and that the theoretical error values are too small by a factor or 2-3. You are talking about only the random part I guess. Do you have an estimation of the systematic error budget as well (spectroscopy, …)? Providing here in Sec. 3.1 a “theoretical” random and systematic uncertainty budget for individual PAN retrievals would be helpful for all the users (in addition to the estimated value for the precision provided in this study by the std of the comparisons).

P5, l-130-135: Do I understand well that PAN is retrieved in the Red windows, but that the previous steps (temperature, H2O, …) are fitted in the large window 760-860 cm^-1? Because I guess fixing H2O is important if the strong H2O line is not included in the PAN fit. Maybe clarify.

Fig.5: It’s nice to see what is measured by CrIS within one day. I guess OE retrievals take too much time to obtain global seasonal maps, and so you have to focus on short periods and collocated regions (such as for these aircrafts campaigns)? It would have been nice to see seasonal global distribution of PAN from CrIS: how long would it take to process, for e.g. one year of data?

-Section 3.2 Initial guess and a priori constraints: “The initial guess profile values for these CrIS PAN retrievals are set to a vanishingly small number.”

 Do I understand from that sentence that the a priori profile (used in Eq. 1) are different than the initial guess profiles? If yes, why not using a priori profiles as initial guess?

- Section 3.2 (should be 3.3) Vertical sensitivity

Could you give the mean DOFS and std for, for example, a typical day as in Fig. 5.  And if the std is large, a little bit more information on which conditions give the best DOFS (e.g. enhanced PAN values, but maybe other factors are influencing the DOFS)? Also you could provide the mean and std DOFS for the PAN retrievals used in the comparisons, which are reflecting more “background conditions”. Also, how much do you lose in % of DOFS by taking 800-300hPa information instead of total columns?

Section 4 Results

P7, l. 198-199: “(most of the aircraft measurements are at the low end of the range)”: What do you mean? Not clear for me.

P7, l. 203-205: The correlation with a priori is already good. Especially with the GT_CIMS, where the improvement with retrieved data could appear limited (0.64 compared to 0.53 for a priori). But looking at Fig. 7, it might be due mainly to the isolated point at 0.46 ppbv. Maybe use robust statistics to derive the correlation and the slope and reduce the effect of a single point (it might reduce correlation with the a priori while keeping good correlation for retrieved values and then strengthen the added value of your retrievals). The slope with retrieved values would also be more accurate by using robust statistics.

Section 5 (“Discussion and Conclusion”)

P8, l. 234-235: While the comparisons are made very carefully and the conclusion “The results … demonstrate the ability of the CrIS PAN retrievals to capture variation in the “background” PAN values observed over remote ocean regions from Atom” is certainly reached, I would be less assertive concerning the statement “Based on this study, we expect this bias to apply to all parts of the world”.  

Indeed, while we see in Fig. 2, that the aircraft measurements often sample PAN levels up to 0.48 ppb, the comparisons with CrIS is limited to 0.32-0.34 ppb (with the exception of only one coincidence), I guess due to collocation. When we look at Fig. 5, where the PAN retrievals can reach 0.90 ppb in some regions, I think that the present study is not covering all the gradient of PAN concentrations to make this statement. Validation at high concentration conditions should be done before concluding that the bias would be the same. It is quite usual to have different satellite biases over clean and over high concentration sites (e.g. TROPOMI HCHO, Vigouroux et al., 2020; TROPOMI tropospheric NO2 Verhoelst et al., 2021; …). Of course, with the interesting approach used here (validating 800-300 hPa, and not tropospheric column), the results might be different but I think it’s worth looking at additional validation at high concentration conditions before concluding that the bias will apply to all parts of the world.

Minor or technical comments:

References Payne et al, 2014, 2017 are missing.

Figure 1, legend: NEdT is not defined

GENERAL COMMENTS

================

The paper describes a new level 2 data set for the CrIS instrument. The retrieval

process is briefly described and an in depth comparison to in situ data

gathered by the ATom campaign(s) is given. Due to the number of CrIS instruments

in orbit and in planning, this is an important data product.

The paper identifies a strong bias of the provided data, which is larger than

the assumed uncertainty. A H2O-VMR-based bias correction is suggested in the User Data

documentation, but not discussed in the paper itself.

The paper should address the bias more explicitely and discuss causes and corrections.

Ideally, the root causes for the bias should be identified and the data product

improved.

I recommend publication after revising the paper to discuss these points in detail and

answering the other comments below.

MAJOR COMMENTS

==============

line 132

--------

The paper identifies a bias of -100 pptv in the derived data, which is larger than

the supplied uncertainty in the data (80 pptv) derived from the standard deviation

computed from differences to in situ measurements.

This suggests that the bias is real and significant, particularly for non-polluted

airmasses outside of plumes.

The employed spectral region is full of emission signatures of a wide range of trace

gasses. It seems as, e.g., CCl4 could still have an effect, but also other CFCs, or

ClONO2 emit in this region. While the strong H2O emission line at 785 has been

avoided, weaker lines are certainly present in the left window.

The User Guide for the data even provides a bias correction formula depending on

water vapour.

I question the usefulness of the data set in the current state.

1) Why was the obvious and *astonishingly* stable bias not corrected in the data set?

2) Why was the retrieval not improved to the point, where no bias correction is necessary?

3) Why was the bias correction formula of the User guide not mentioned or applied for the comparison?

These points need to be adressed within the paper.

Figure 2

--------

High PAN VMRs occur often at higher tropospheric altitudes (particularly due to the longer

lifetime at colder temperatures) close to the tropopause. The used aircraft data

rarely go above 12km.

Biomass burning plumes reach higher than 12km, particularly in the tropics.

The given altitude range of 800hPa to 300hPa is key here, as 300hPa

corresponds roughly to 10km.

How does this limited altitude range affect the accuracy of estimating total PAN in the UTLS?

Why is the instrument not sensititve (at all? enough?) to high PAN VMRs closer

to the tropopause? Is this related to the low temperature at this altitudes?

SPECIFIC COMMENTS

=================

line 135

--------

Particularly in the face of the discoved systematic error, a discussion on the sensitivity

of the retrieved PAN VMRs on the previously derived quantities (i.e. the 'b' vector) might be

interesting. It is mentioned that the retrieval processor under-estimates the "observation error",

without detailing what exactly this entails. Often this only contains - for practical reasons -

an estimate of the noise induced error, not the systematic errors.

How does the identified systematic bias relate to the error diagnostics for systematic (b-related)

errors?

line 165

--------

Please show a (representative set of) averaging kernels to show the region of sensitivity.

Figure 4

--------

The residual shows structure beyond the noise level (blue lines). The caption indicates that

this spectrum was computed with a zero PAN profile. Please show both a spectrum with the

derived PAN profile and with a zero profile to show the improvement and PAN signal as well

as quality of fit of the used spectra (similar to Glatthor et al., 2007)

Figure 5

--------

The paper identifies a low bias of 100 pptv causing many VMRs to be negative as shown in

Fig. 7. Figure 5 shows only positive VMRs. Please explain the discrepancy.

MINOR REMARKS

=============

line 108

--------

A big X with a hat was not in (1). Maybe big-hat-x -> hat-x and hat-x-a -> hat_x ?

line 113

--------

\Delta f should be 'bold'.

line 128

--------

an approximate solution?

line 134

--------

CCl_4 (small l)

line 139

--------

It is not clear from the context what the "forward stream" is. The given reference

distinguishes a "reanalysis stream" without being clear on the difference.

I suppose it has something to do with using (forward) extrapolation of calibration data in

contrast to interpolation using (later) data. This is probably a very common term in certain

scientific communities.

Maybe explain it in a brief sentence.