This article introduces a deep learning model called ANTSIC-UNet for predicting the extended seasonal variations in Antarctic sea ice concentration, with the ability to forecast up to 6 months in advance. The study utilized a rich set of climate variables for model training and compared it against two benchmark models (linear trend and anomaly persistence models). The results demonstrate that ANTSIC-UNet exhibits superior predictive skills in sea ice concentration and integrated ice-edge error, especially in forecasting extreme events in recent years. The strengths of the article include the consideration of both sea ice and related atmospheric and oceanic variables enhances the accuracy of the predictions. The results are interesting and the work could be published after moderate revision. My comments are intended to improve the presentation of the paper and require clarifying unclear points.

Comments

1. L76“57 is the dimension of the variables”However, when we calculate 12+1+14*3+1, it equals 56. So, what is the extra one?
2. L184 For September, compared to anomaly persistence, ANTSIC-UNet shows a larger negative bias in the sea ice edge region. What could be the possible reasons for this error?
3. L186 Is the lower RMSE in September compared to February related to the size of the area considered during the calculation? Are the regions used for calculating each indicator consistent with the respective months?
4. L274 Is the high importance of variables in the model due to the seasonal cycle? Does the importance of variables change for SIC anomaly?
5. The main improvement of this article compared to other DL methods is the inclusion of relevant variables that affect sea ice in the training data of the model. How significant is the impact of these variables compared to a model trained solely using historical data？
6. The section on the importance of each variable is very insightful. The author presents some viewpoints that are inconsistent with statistical models, such as the minimal impact of variables like temperature and wind speed in DL methods. Does this suggest that DL methods have not learned the underlying mechanisms of these variables to some extent？

This paper documents the results from a deep learning effort at predicting maps of Antarctic sea ice from the NSIDC. The model is generally well-described, with well documented results that are effectively compared to simple linear trends and anomaly persistence. However, the paper focus is only on the performance of a single effort of sea ice prediction and contains no effort to use this tool to add any scientific knowledge or insight to Cryospheric science. There is only the briefest attempt to contextualise the importance of Antarctic Sea Ice prediction, and the reasoning behind the variable selection is not described at all. There is very little documentation on how the model was developed and any insight into what was learnt during the development process. The publishing criteria for the Cryosphere is that there needs to be a scientific aspect to the publications beyond model description and results, therefore this paper is not acceptable=and in opinion needs to be rejected.

Particular issues:

Throughout the paper there is a lack of knowledge of the system that is being investigated, and the study is only focused or representing the input data and the physical system. For example the title says it ‘predicts’ sea ice – there are many aspects of sea ice that are not considered here. This paper only looks at monthly sea ice concentration maps from the NSIDC – possibly the simplest representation of sea ice. There are many other datasets available – this needs to be documented. The introduction is very brief and contains no description of the system being investigated.

The most useful aspect of the study can be to inform of what variables from the chosen reanalysis are the strongest predictors. This is attempted in section 3.4 – but it has no contextualization. Key aspects that need including: Why physically may each variable be useful in prediction? How accurate are each variable within the reanalysis product? The lack of predictive importance for Downward solar for example may be due to this variable being poorly represented within the reanalysis. What other scientific analysis has been performed using this reanalysis? How has it been used outside of Deep learning to investigate sea ice?

Finally there is little to no contextualization of results amongst contemporary literature and other prediction efforts. Section 4 contains only a handful of citations when it is essential to contrast the results here with other efforts at sea ice predictions. How do the reported skills in forecasting compare to other efforts, Andersson et al. (2021) is an important bench mark here. How are the extreme years (2017, 2022, 2013) described in literature? What other hypothesis exist about what affected sea ice in these years?

“sea ice concentration” or area or extent needs mentioning in the title.

L 9 the changes to Antarctic Sea Ice a subtle and require more than this introductory sentence – after a period of increasing summer minima there have then been reductions.

The Abstract needs more description on why Antarctic sea ice needs predicting.  L 15 – 20 can be removed as this is too much detail for an abstract. The final 5 lines are ok as a summary. Some contextualization amongst previous publications is needed for the abstract too.

L 25, a first general sentence on the nature of sea ice will help here.

L 27 -  this is only true for the summer minimum.

L 29, variability in what? I guess extent?

L 32, everything here after “like” is very vague and needs rewriting.

L 37 these are-ice-sea interaction processes need further description.

L 64 this sentence is difficult to follow. Are linear monthly trends extrapolated to future dates used as a model input?

L 66 a description of why reanalysis data is sought is required either here or in an earlier description of the project incentives. What do each data represent and why are they needed for predictions?

L 76 Is the the input data volume held static throughout all development? The data lag is often an option that requires testing and investigation.

L 72  why is v10hPa not included also?

L 108 The linear trend prediction is not described well in section 2.1

L 112 This implies that the RHS of equation 1 is just the observed ice concentration field. What benefit is this? Further description of how anomaly persistence works as a prediction is needed here.

L 165 – key acronyms need defining in each figure caption. ( and all others too)

L 253 “extremely low” rephrase with better accuracy.

L 253 -  this table needs extra columns to show what was extreme about these years – SIE/SIC anomalies perhaps.