A 20-Year Climatology of Sea Ice Leads Detected in Infrared Satellite Imagery Using a Convolutional Neural Network
Round 1
Reviewer 1 Report
The paper describes the 20-years (2002-2022) climatology of MODIS-detected sea ice leads in the Arctic with an AI approach.
The paper is interesting, but some aspects relevant to possible biases which may arise from the analyzed data have to be carefully considered.
In the following I report my comments:
On p.2, row 75 the authors state that VIIRS data were not included in the analysis "... because the increase in observations has resulted in more lead detections." My questions are: In which sense the "more lead detection" is a biased result and what are the reasons causing the bias? Does it mean that the AI technique described in the paper can be employed only if 1 single sensor is used? If so, this implies that an optimal temporal observation rate could be defined to guarantee "unbiased" lead detection. Do MODIS observation rates accomplish this requirement?
In my opinion, the use of multiple sensors, especially at different hours, could indeed improve the rate of detection, for example in the case that if one observation is obscured by cloud cover, the next one may result successfully if, meanwhile, the cloud cover disappeared. At the very least, in case the same lead is detected twice, it can be easily accounted for (I guess).
The second main comment is relevant to eq. (1), which expresses the linear relationship between the "clear sky frequency" and "lead frequency".
The fitting result states that if the clear frequency is zero (i.e., 100% cloud cover for all times), the detection of leads is negative, specifically -2.391 in value. That is absurd as, by definition, the lead frequency is always a positive number with a well-defined physical meaning. This means that the fitting law may be wrong, not so much about the assumed linear trend, but rather that the straight line should be constrained to pass through the point (0,0). So, I guess the extrapolated results discussed throughout the paper include a bias, whose amount should be quantified.
Finally, I would suggest the authors include a visual representation of the effective spread of the "lead area" values. As an example, this may be achieved using a box-plot format.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
excellent article, no changes needed
Author Response
Thank you for your review.
Reviewer 3 Report
This manuscript, as written, is trying to answer two questions: (1) are the leads in ice decreasing or increasing in time? and (2) is the AI method more suitable that other used methods?
The authors find as a result of their analysis of remote sensing data, that detection level of leads decreases in time, and that they suggest that the leads may really be increasing in time and that the decrease in detection is due to changing (increasing) cloudiness. Thus, considering geophysical significance of research there is nothing the paper can say. The leads are either increasing or decreasing in time depending how you factor in other influences.
The authors conclude that “… leads are not becoming less likely to occur, and are likely increasing at a rate of 3,700 ±3,500 km2/year.” The rate of increase is about the same as uncertainty of the result. This is very weak conclusion based on an extensive description of work performed. This sounds as a negative result. I believe that even negative results should be published, so that other do not repeat the same useless work.
Another question is the method used. The authors conclude that the AI method is consistent with or better than the earlier method. It looks like they are hesitant to say clearly that the AI method perform better. I think you should say clearly what is your assessment. It is possible that the AI method is not better than older method, but that it has some technical advantages in its application.
Although I have no objection concerning technical procedures, I do not recommend publication in the present from. I urge the authors to state clearly and briefly the results even if they are negative. If you can describe something in one paragraph, don’t write two. In conclusion, the significance and impact of described research is low.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have convincingly replied to my comments. The paper has been improved by clarifying the raised points. Therefore, I recommend publication in the present form.
Author Response
Thank you for your review.
Reviewer 3 Report
No objection.
Author Response
Thank you for your review.
