Review of egusphere-2023-1297

"MMSeaIce: Multi-task Mapping of Sea Ice Parameters from AI4Arctic Sea Ice Challenge Dataset"

General Comments:

• This article reads extremely well, with no issues with the English language use and with good clear content.
• I only suggest that the scientific novelties are presented with more emphasis than the competition for this scientific paper, rather than as an accidental consequence. The novel learnings and explanations are what makes this a scientific work, rather than just your method documentation. That said, there is far more novel content here than many recent ML submissions.
• I conclude that this work is valuable and worthy, but should be revised to emphasise the scientific messages.

Specific Comments:

• The mentioned imbalance of competition and scientific novelty is clear in the Abstract. The real science only appears in the very last sentence, with the showcasing of the various techniques, or components of the system, yet this is where the real scientific advancement lies. The abstract, and the rest of the paper, should summarise these messages and what we learned from the exercise.
• In terms of science, the method should also explain why you designed the network as you did, and why you designed the ablation study as you did?  How does it characterise the significance of the different components? Were the components included or developed with certain expectations, e.g., have they been used before in different contexts perhaps? This is where we can learn the most about your method and the importance of various components.
• Consider whether the title can somehow reflect that the science is somehow this contribution/significance analysis of the components. Might be difficult and is not critical though.

Technical Comments:

• I suggest that you add sub-headings on the left with the different models to explain what they are and their relation to the ablation study and table 4. That is, remind the viewer which is the "full model", that model 2 has "no downscaling", and model 8 uses "cross-entropy", etc. This would make it easier to try to interpret the causes of the results.

First of all, I would like to congratulate the authors on their first place in the competition and a well-written, concise and informative report of their findings.

Broad Comments:

From a technical standpoint I find the manuscript to be well constructed and easy to follow. I do believe some extra discussion would benefit the work and help place it into the greater context of the ongoing efforts of sea ice classification in a changing Arctic.

The two things I would like to see discussed in additional detail would be:

1. The influence of the ice charts as ground truth in terms of what the resulting classifier is capable of extracting and what is outside of the scope of classification. In the introduction some of the uses of ice charts and the multitude of output variables is mentioned; it seems to me that a regional sea ice concentration and floe size as is predicted here, could be derived from a classified map if the classification took place at the same effective resolution as the SAR sensor, for example.
2. The effect of including time and spatial information in the classification and what that might mean for using such a classifier in a changing Arctic. In a wider scope, one could ask the question if there might be a conflict between performing best on historical data and performing best in an uncertain future. This can be discussed in terms of which input variables are used, how the class imbalance is handled, etc.

Specific Comments:

L.10: The authors claim that the tested techniques significantly improve the robustness of models, is this a qualitative finding or is there some quantitative analysis backing up this statement? Maybe this is unclear because robustness is not uniquely defined in this context.

Sec 3: I am sorry if I just missed it, but I would like some discussion on input data preparation. I assume that some of the auxiliary data was brought up to input patch dimensions and added as channels because of convenience, but might this have an effect on the classifier (e.g. vs adding them in the bottleneck)?

L.129: Why were the months discretized for input instead of a continuous approach and what are the possible implications for the classification?

L.197: The predictions aren’t really ‘polygon based’ are they? Maybe spatially smoothed predictions or some similar wording might be more fitting.

L.197-200: Some published methods exist that make use of various input scales, maybe this could be mentioned/referenced here.

Thank you for considering my comments.

See attachments.