Review of Shortening of the Arctic cold air outbreak season detected by a phenomenological machine learning approach:

The authors have produced a novel method of characterizing cold air outbreaks (CAO) in the north Atlantic using an unsupervised automated routine and infrared MODIS imagery. The product is compared to a few established thresholds using reanalysis data, and is shown to be more accurate. Cold air outbreaks are shown to be more common in winter and shoulder seasons compared to the warmer parts of the year, and CAOs are increasing in winter, but decreasing in spring and autumn. Most of this signal is shown to be occurring in the southern portion of the study region.

The novel CAO product is clearly superior to existing measures of CAO, so this publication and product is relevant to the community and certainly worthy of publication in ACP. The manuscript presents and describes the product well, but the trend analysis and attribution portions require some additional work to clarify and solidify some of the propose mechanisms. I recommend that this article be accepted after some major revisions.

Main points:

1) The discussion of processes concerning open cells may be lacking. Particularly on lines 37-38 when open cells are tied to drying processes. This is somewhat contradicted by Eastman et al., (2022) that shows the closed-to-open Sc transition associated with increased boundary layer moisture and stronger fluxes (particularly surface winds and precipitation), in contrast to the transition to more disorganized cloud types, which are associated with drying. Increased SST is likely associated with stronger fluxes, so an SST-driven mechanism is still probable here, but the mechanics are unlikely due to drying processes.

Reference: Eastman, R., McCoy, I. L., & Wood,R. (2022). Wind, rain, and the closed to open cell transition in subtropical marine stratocumulus. Journal of Geophysical Research: Atmospheres, 127, e2022JD036795. https://doi.org/10.1029/2022JD036795

2) Throughout the manuscript, there is discussion of trends specifically for open cells (for example, line 485). This seems a bit speculative, since the CAO dataset does not appear to directly assess cellular structure, unless I am misinterpreting something. It may be wise to tone down these assumptions, since changes in MCC structure aren’t really being assessed here.

3) Line 390: I don’t see a 20% increase in CAOnet in March, but I do see one in December.

Line 395: I also don’t see any significant decreasing trends for any winter month for any of the CAO measures.

It is possible that I do not have a current version of this figure in my manuscript, but the discussion of Figure 7 does not appear to line up with what I’m seeing on the figure.

4) As far as attribution of trends is concerned, much of this appears entirely speculative. In fact, the authors do not show any correlation analysis between time series of their CAO data and SST or ice edge data, which may strongly aid any discussion of attribution. I recommend revamping this section in order to more clearly show these proposed relationships using your new CAOnet data, or changing it a bit to motivate future work to do this, while not attempting to attribute the trends to anything here. Additionally, are any other flux variables known to be changing in this region? Trends in wind speeds or direction may be interesting.

5) Concerning the trends: The method that is being used has been called the ‘median of pairwise slopes’ method, and it may improve discussion of that approach if you used that name, since it is fairly intuitive.

Further, it would really aid the work to do a little more to characterize the trends. A split between CAO frequency (how many days are CAOs detected, regardless of size), and amount when present (how much area is covered by CAO clouds) may help show whether CAO events are changing in frequency or size, or both.

6) Finally, it would strongly benefit the discussion of radiative characteristics to compare cloud amount observed when a CAO is occurring to cloud amount observed when it is not occurring. This is hinted at in the text, but is an essential result in order to actually add any value to a discussion of radiative impacts of CAOs.

This work presents a new, machine learning-based method of identifying clouds associated with cold air outbreaks, seeking to overcome biases which may arise from using environmental conditions to predict the occurrence of such clouds. The existence of such a database would be useful to the community. Overall, the study is well-written, the methods are generally well-described for an audience which is not familiar with ML, but overall could do with describing how it fits in the context of existing work.

However, I have some questions regarding the methods, and the manner in which they are compared with existing approaches, addressed below. I also recommend some further quantitative analysis to support some statements on the mechanisms underlying the identified trends, and the radiative impacts of these clouds. Overall, I recommend for publication, following some major revisions and additional analysis.

Points to address:

L23: Given it’s the Arctic, perhaps say the seasonal radiative effect of CAO – they don’t have a strong cooling effect in the winter, etc. This is a strong motivation for the study, so warrants more discussion of existing literature.

L51: Can you explain more why the positive index is an issue? I understand the uncertainties in the analysis data, which is very good motivation. But, for example, studies which track the clouds (etc., Murray-Watson et al., 2023), they specifically identify clouds moving from ice to ocean with high indices, thus potentially limiting the false identification of the CAO clouds.

Section 2.1.2: Variability in referencing: He et al is given as a reference for residual blocks improving performance, but nothing is given for the leaky ReLU. This might be convention for the ML community, though.

Section 2.13: Why only five years of to train data if 25 were available? I thought usually the test/train split was 80:20, on a whole dataset? And are 500 swaths a robust evaluation dataset, if 15,200 swaths are available? Perhaps I am misunderstanding what warrants a statistically significant evaluation of ML techniques.

Line235: Are they weaker CAOs, or CAOs further along in their development?

Lin262: Is there any way to rationalise why there are two CAO clusters? Is it early/late in development? What characteristics split them?

Section3.1: How do you account for overlying clouds affecting the M-based classification in the imagery, which you’re automatically filtering out of the CAOnet one? Is it a fair comparison?

Line275: The high M index probably doesn’t capture the open cells well because they are downstream of the CAO development, no? So are they not still CAO clouds, just advected downstream? Tracking studies seem to show that M decreases downwind of the sea ice edge, so using this gridbox-by-gridbox metric with M > 3.75 will necessarily miss these clouds. This may also affect later discussion about Climatology, etc.

3.3: I’m not sure some of the figures mentioned here align with what Figure 7 shows. Also, broadly, I think this section could do with some more mechanistic analysis. They discuss the literature, but it would be helpful if they could show how their tools could be used to answer these questions. It is more speculative than explanatory.

L525: Similarly, the conclusions state that the trends are driven by faster warming of the air compared to the sea, but this isn’t necessarily shown.

Similarly, considering the radiative impact of CAOs in a changing Arctic was a strong motivation for the study, some analysis would be appreciated. Conclusions are inferred from changes in trends etc., but again, this is more speculative.