23 These L3L L3W differences are clearly linked to retrieval sensitivity differences

While reading the abstract it was unclear if the authors were taking into consideration the fact that CO emissions over water are negligible and whether that affects the difference between L3L and L3W.

421 As expected from the previous analysis, the land-water sensitivity contrast is greater when mean VMRs are significantly different than when not.

It’s as if they assume that all land-water contrast is a processing artifact. See line 446

446 An underlying assumption is that the temporal trend in “true” VMRs should not vary much across a 1o x 1o L3 grid box.

This was revisited in line 399. It seems as if the authors neglect to account for wind direction. If the prevailing wind is blowing from the ocean inland (e.g. the coast of California) then the CO concentration could be much higher over the land than the water in the same gridbox. Whereas if the wind was blowing out to sea, one would expect far less difference between L3L and L3W. Yet it appears that the authors are not accounting for this.

110 It cannot be overlooked that working with L3 data thus requires fewer computing resources and less technical proficiency

Agreed. Furthermore gridded products are accessible by many more tools that users are familiar with such as Panoply.

177 Validation results are comparable to V8. It is expected that the main conclusions of this paper to hold for V9, since the land-water sensitivity contrast remains and L3 processing method appears to be unchanged.

Actually, V9 discards far fewer L2 pixels due to cloudiness than V8 which may affect the results. I suggest the authors repeat some experiments with V9 to confirm their assumption. Yes, they are correct that the L2 → L3 processing method is unchanged.

193 which at the time of writing, is the most recent data quality summary

More recent data quality statements are available now. See https://asdc.larc.nasa.gov/documents/mopitt/mopitt_quality_statements.html

483 However, the results presented do imply a general tendency for trend

484 underestimation in retrievals over water within coastal grid boxes compared to retrievals over land in the

485 same grid boxes obtained at the same times, which appears to be linked to differences in retrieval sensitivity.

This feels like the most important point of the paper. Perhaps more effort to demonstrate and quantify would be helpful.

The discussion in the paragraph starting on line 417 is very important however it would have even more impact if it it included the consideration of one more bit of information. The skill of MOPITT retrievals of CO (VMR) is not random. It is dependent on conditions such as thermal contrast between the surface and the air, which is what the authors are describing when they see discontinuities between L3L and L3W. However another factor is that MOPITT sees CO better when there’s a lot of it. The uncertainty (as measured by DFS) decreases when the CO signal is large. So if there’s less CO over the ocean due to fewer sources, that will also affect the results of this analysis.

451 – The word “gradient” appears but I’m having trouble understanding its definition here. Is it just the difference between temporal trends of L3L and L3W? Or is it spatial?

Table 1 – It took me several attempts to understand what the “d” column was.

483 However, the results presented do imply a general tendency for trend

484 underestimation in retrievals over water within coastal grid boxes compared to retrievals over land in the

485 same grid boxes obtained at the same times, which appears to be linked to differences in retrieval sensitivity.

This seems like a valid conclusion based on the analysis performed. There are a lot of details about the methodology to arrive at this conclusion that confused me more than served as support for clear statements such as this.

613 In these instances, L3O would therefore seem to be misclassified.

This is a valuable insight.

820 there is enough

821 evidence to support the suggestion from Ashpole and Wiacek (2020) that an additional L3 “land-only”

822 product, created only from averaging bounded L2 retrievals performed over land – the L3L dataset that has

823 been analysed in this paper – would be beneficial to the research community.

This recommendation will be brought to the attention of the MOPITT science team. It will hopefully be incorporated in the archival processing version.

General Comments:

These researchers took a very close look at how the MOPITT L3 product is created and have identified a flaw in the way the MOPITT team processes pixels into coastal grid cells by mixing retrievals of uneven quality. This distorts the values reported for a non-insignificant number of gridcells. Their conclusions appear valid and robust. However, I had a difficult time following the arguments and methodolgy of the paper. I didn’t understand why they were focused on surface level retrievals instead of higher in the atmosphere where MOPITT is considerably more sensitive. I was curious if they would have come to the same conclusion if they looked at MOP03M (monthly mean) products which have far less random noise and greater coverage than the daily L3O products. In several places, the authors were making a clear distinction between two situations and I had trouble understanding the meaning of this distinctions. For example: "For other datasets, whether the marker is filled or not, and whether the lines are solid or dash/dot, depends on the outcome of an independent, 2-tailed t-test assuming unequal variance (aka “Welch’s test”) against L3L: filled markers and solid lines indicate the mean is significantly different to L3L (p < 0.1); open markers and dash/dot lines indicate there is no significant difference to L3L." This distinction was too difficult for me to understand its significance.

I believe the researchers can transform this paper into a valuable analysis by having a more clarified statement of their conclusions and focusing the readers' attention on the evidence that supports that point.

The manuscript describes a study of MOPITT V8 TIR-NIR surface CO retrievals over 33 coastal cities. Daily L3 data (data gridded to 1^ox1^o, 111x111 km² per pixel) and daily L2 data (22x22km² per pixel at nadir) are analyzed. This study’s main findings are that statistics of coastal cities obtained from L3 and L2 products differ, that "mixed" L3 pixels (L3 pixels averaging both water and land L2 pixels) are not suitable to study coastal cities, and that a L3 land only product for coastal pixels is needed. In order to demonstrate these points, several comparisons and statistical analyses between land and water L3 TIR-NIR pixels (original and re-created from L2 data) are performed. The manuscript is well written.

Two major issues are described below.

1. Use of TIR-NIR data in land/water comparisons

As described in Deeter et al., 2013, among others, TIR-NIR retrievals over land and over water are fundamentally different, since NIR radiances cannot be used in the latter. The authors acknowledge the fact that retrievals over water are limited to the TIR band due to the lack of NIR signal, but don’t acknowledge the implications, which are key. Using the TIR-NIR product for this study is not appropriate, since there are two effects causing land/water differences in the averaging kernels: thermal contrast effects and the lack of NIR radiances in retrievals over water. The two effects cannot be separated.

2. Use of L3 data to study coastal cities

L3 products (either TIR-NIR or TIR) are not suited for the analysis of the coastal cities listed, given the horizontal extent of the targets. A cursory search (please see Table 1 attached) shows that 30 of the 33 cities in Fig. 9 correspond to a very small fraction of a single L3 pixel footprint. Only 3 of the 33 cities are close to covering or barely cover one L3 pixel footprint. Basing such analysis on L2 data could be an adequate choice, at least for some of these cities. About half of the 33 cities would not even fill the footprint of a single L2 retrieval. Only 10 of the 33 cities would fill 4 or more L2 retrieval footprints.

According to the manuscript, “L3 data are better suited to long timeseries analysis than L2 data owing to their smaller size”. That statement is wrong. Some tools are easier and more convenient to use than others, but that does not mean that they are better suited for a given task. Analyzing long time series with L3 data may be easier, more convenient. However, easy and convenient generally comes at a cost, in this case the quality of the analysis. The manuscript continues “working with L3 data [...] requires fewer computing resources and less technical proficiency […] L3 products thus make the MOPITT data more easily accessible, especially to less-expert users, who may lack the expertise required to scrutinize the data for potential a priori bias.” Again, a tool may be easy/convenient to use but unfit for certain tasks.

Time series are at the center of this work and are the justification provided for using L3 data in the first place. The manuscript, however, does not include a single time series. It’s hard to imagine that meaningful information/trends can be identified in L3 time series covering a ~6400 days range (from 25 Aug 2001 to 28 Feb 2019) but having only a few hundreds of even a few tens of days with a L3 value at all (and that L3 value coming in all cases from a single L3 pixel). This is the case for most of the cities analyzed (Fig. 9).

Are those few hundreds of even few tens of L3 data points representative of the 1^ox1^o areas they stand for? L3 pixels (land, water, or mixed) may be produced by averaging as little as 2 L2 pixels. As an example: more than 25% of the total number of daytime pixels in a randomly selected L3 file resulted from averaging either 2 or 3 L2 measurements. These L3 pixels may not be representative of the 1^ox1^o area they stand for and, thus, should be filtered out so as not to corrupt the statistical results. It is unclear if such filtering was applied.