Dear authors,

I have attached my comments in a document. I would consider major revision, but it is just considering my point about the trends which is an open question and that could lead to some modifications of the manuscript.

Do not hesitate to contact me if you have questions,

All the best.

Marion

Review of “A long-term proxy for sea ice thickness in the Canadian Arctic: 1996-2020” by Glissenaar et al.

Summary

This paper presents a 25-year sea ice thickness proxy-product for the Canadian Arctic, showing thinning over most areas in the Canadian Arctic, with variations across regions and months. The proxy-product is based on a Random Forest Regression model trained on CryoSat-2 observations, estimating thickness within 30 to 50 cm uncertainty. Validation is carried out with independent measurements from moorings and NASA’s Operation IceBridge (OIB) as well as in-situ thickness records from three locations in the Canadian Arctic. The Beaufort Sea and Baffin Bay show significant negative trends throughout the year, while the Arctic Ocean Periphery shows thinning above 2 cm/yr during all months but April, with a peak of -3.3 cm/yr in December. The Parry Channel shows weaker thinning trends, but with high yearly variability. This proxy-product provides a long-term record for studying trends and variability in sea ice thickness in the Canadian Arctic, including the narrow channels in the Canadian Arctic Archipelago.

General Comments 

I think this is a creative approach of how to extend sea-ice thickness time series, especially in regions, that are difficult to access with satellite altimetry data (close to the coast). I appreciate the idea of synergizing different types of information (different satellite sensors, ice charts) to improve our knowledge of sea ice thickness, especially where altimeter data leave gaps.

The paper is easy to read, but I find there is a lack of information regarding the data and methods, especially scatterometer data, but also ice charts, partly "hidden" in the supplements. For example, I recommend being more specific on the used ASCAT data. Which data product have you used? The sigma-0 at 40 deg incidence? Moreover, information like the workflow diagram is key to understand the study and should be included in the paper. In general, I think information, crucial for the paper should be present in the actual paper, while Supplements only support the paper, e.g. additional plots that show something in more detail. Some other occasions where I think clarification or additional information is needed, you will find under "specific comments".

In addition, I have some major scientific comments/questions:

1) I wonder if the output can be improved by using different training data. The Beaufort Sea is known for being very challenging for satellite altimetry, because of the mixed ice types, low SIT correlation lengths, and sometimes high drift speeds. I suppose that this leads to some misfit between altimetry, scatterometry, and ice chart data. What about using only areas, where confidence in CS2 SIT is better? There are also CS2 SIT products that provide data in the Canadian Archipelago, for example the AWI product. It would be interesting to compare the proxy product with such estimates, too. 

2) I would expect a certain bias between scatterometry data sets from different sensors (but same frequency bands). For example, between ASCAT and ERS, because the sensors are different (as pointed out in the supplements). Can you rule out a bias that might affect derived trends in the proxy product? Or should it be at leased discussed in the limitations section? 

3) A major concern is the correction to thinning of ice types (3.4). I understand the problem, but PIOMAS also comes with considerable uncertainties in the study regions. And it is a model, too. I wonder if it makes sense to correct the proxy product (based on a ML model) with trends from a sea ice model, while one could then also just take PIOMAS to look at trends. Figure 11 actually shows (except Baffin Bay may be) that the trend in the corrected product is basically coming from PIOMAS? I think it is feasible to compare with PIOMAS trends to discuss the problem. But I am not sure if this can be sold it as a separate product. I would argue that it is less confusing for potential users if there is just one product, where limitations and uncertainties are clear.      

4) The uncertainty estimate of 30-50 cm is based on the estimated model uncertainty, verified by the ULS comparison, if I understand correct. How do the OIB data compare to the proxy product in numbers? Section 4.2.2 is rather descriptive. It would be good if some numbers to verify the uncertainties can be presented here as well, like RMSD values etc.   

Therefore, I suggest major revisions. I think the idea is very good and worth to be published. But the concerns above should be addressed.

Specific Comments:

L81: I think the justification should be formulated here.

L122: Perhaps mention that scikit-learn is a python library.

L123: Any reasoning why 95?, and the value for the maximum depth? Is it an empirical choice, after trying different setups? 

L136: May be consider introducing sub-sections for each validation data set.

L147: Is it the same ice density as used for the CS2 SIT product? Why do you not distinguish between FYI and MYI density, see Alexandrov et al. (2010), Jutila et al. (2021)?

L149: I think it should be clarified that OIB does not measure SIT but uses the ATM laser and the snow radar to measure snow freeboard and snow depth and convert freeboard into thickness. This conversion goes along with several uncertainties as well. So, I suggest to rather write that OIB provides SIT "retrievals" or something similar. 

L178: I suggest shortening the lower case terms here to improve readability of this formula. And what is the difference between the "i" and "icecategory"? Please clarify.

L191: May be shortly explain the "10-fold cross validation RMSE", what does it mean? And is the testing error directly related to the uncertainty given in the Abstract (30 to 50 cm)?  

L199: How have you chosen the 20% CS2 data? Are they randomly picked points? Or did you cut out a certain area? This distinction can be important as the correlation between both chunks of CS2 data might be different. Please clarify in the text. 

L217: The usage of "proxy product" and "not corrected proxy product" is sometimes confusing. May be use a more consistent nomenclature, e.g. proxy_corr product and proxy_nocorr (or only proxy) product, throughout the paper. 

Figure 2: I do not understand Fig. 2b) - what does "form of ice" mean here? And why does it have "km" as unit? Shall it relate to floe size? Or shall it show different forms of ice more in the sense of ice type, but then the colormap should be discrete?

Figure 8: The histograms are very difficult to separate. Maybe you can have the columns next to each other (like in the supplements) and/or use colors that are easier to separate.

Figure 9: It is very difficult to compare the in-situ values, as the time axis is very coarse, while the interannual gradient is quite strong.   

Figure 11/12: For the line plots, it would help to have legends. Why are you using colors for the different regions if you present them in different boxes? I am also slightly confused with the "solid and weaker circles". "Colours show the trend for the not corrected version" -> But they are all colored?  This figure not so easy to read.

Figure S1: I suggest including this figure in the paper, as it shows the workflow to derive the proxy product - too crucial for the Supplements from my point of view.

References:

Alexandrov, V., Sandven, S., Wahlin, J., & Johannessen, O. M. (2010). The relation between sea ice thickness and freeboard in the Arctic. The Cryosphere, 4(3), 373-380.

Jutila, A., Hendricks, S., Ricker, R., Albedyll, L. V., Krumpen, T., & Haas, C. (2021). Retrieval and parametrisation of sea-ice bulk density from airborne multi-sensor measurements. Cryosphere Discussions.