A Suitable Retrieval Algorithm of Arctic Snow Depths with AMSR-2 and Its Application to Sea Ice Thicknesses of Cryosat-2 Data

Round 1

Reviewer 1 Report

All of my comments have been addressed. 

Author Response

Thanks very much.

Author Response File: Author Response.docx

Reviewer 2 Report

The presented manuscript shows and compares several methods for obtaining snow depth from AMSR-2 microwave radiometer. This is important because snow depth over sea ice plays an significant role, but remains poorly studied, so the validation of new methods is not easy task because of the lack of in situ data.

Please, see further questions raised in the general comments:

1. The low correlation coefficient 0.26 between the neural network and OIB snow depths (line 498) is surprising, because in the original paper (Braakmann-Folgmann et al. 2019) the correlation coefficient stated as 0.82. Authors should give the formulas in the text by which they calculate statistical estimates in order to remove possible questions.

2. Equation 1 of snow density relationship (line 289) is better to move immediately after it mention in the text (after line 281). You can also move Figure 2 slightly above, before this text: "The calculation of sea ice freeboard requires us to know…" (line 282).

3. In general, the use of this snow density relationship from passed months since October (Equation 1) is questionable, because this relationship is based on Warren's climatology which data obtained prior to the first significant impacts of climate change, and, as the authors themselves point out above: "this climatology cannot represent the current snow state". This climatology greatly overestimates the currently observed snow depths in Arctic, and consequently, the snow may be less compacted in spring and have a smaller density, than it predicts.

4. Can you expand Figure 1, and give the data that was used for the validation separately?

5. The amount of data to check the model is quite sparse and represents the same data on which the model was trained only for two months of another year. It would be interesting to compare the results with data obtained by other methods, for example:

AWI IceBird data, which since 2019 carried a snow radar: https://www.awi.de/en/science/climate-sciences/sea-ice-physics/projects/ice-bird.html

AWI snow depth buoys data: https://epic.awi.de/id/eprint/40804/1/1604_Nicolaus_SeasonalityOfSnow_small.pdf

Ice mass balance buoy (IMB) data: http://imb-crrel-dartmouth.org/

ASPeCt standardised data: https://www.cen.uni-hamburg.de/en/icdc/data/cryosphere/seaiceparameter-shipobs.html (Dataset: https://cera-www.dkrz.de/WDCC/ui/cerasearch/entry?acronym=SMVSBSIOAV2_3880415 )

It would also be helpful to compare your results with recent altimetric snow depth estimates using bi-frequency Ka/Ku measurements:  https://tc.copernicus.org/articles/15/5483/2021/

6. I think Figure 12 is redundant. In addition, it is quite strange that the results of evaluation of the sea ice freeboard are given after the results on the SIT calculated on its basis are already showed. I think this figure and relevant text can be removed, figure 13 is quite enough.

7. Have you assessed the uncertainty of snow depth estimates with your neural network?

 

Minor comments:

The abbreviation SIC, which first appeared in line 272, should be explained at once.

Several references in a row, such as in line 162 should be written as [43-45] instead of [43] [44] [45]. At line 165 should be written as [32,42] instead of [32][42]. Correct please all such cases throughout the text.

The text “are the” on line 344 does not need to be italicized.

Lines 416 and 417 use a strange font for rho (for water, snow and ice densities)

I accidentally noticed that several DOI in references are incorrect (for example reference 23 and 53). Please check all references for correct DOI.

 

Author Response

The authors would like to thank the editor and the reviewers for your valuable comments to help to improve the manuscript. We have now carefully reviewed and addressed all of comments which we hope meet with approval, with revisions to the manuscript shown in red. Please check word file.

Reviewer 3 Report

This study developed a new snow depth retrieval method over Arctic sea ice with a LSTM deep learning algorithm. The author compared the retrieved snow depth and sea ice thickness based on Cryosat-2 data with the modified W99, 17 AWI, Bremen, Kwok and Neural Network snow depths. Compared to OIB snow thickness data, the LSTM method provides better accuracy in snow depth and sea ice thickness estimation than other methods. In my opinion, the paper is of significant as we are in a new era of sea ice thickness retrieved. The paper was well-written and the authors provided a detailed methodology including the correction, processing steps, etc. The results and discussion were complete and presented in logical order. There are however a few minor corrections to improve the paper quality.

1. Abbreviations of nouns should be written in full when they first appear in the text, such as MYI, RMSE, MAE et al.
2. Equations should be placed under the introduction.
3. Section 2.4: It is advisable to add a table with basic information about several snow depth products to get a better understanding of them.
4. What’s the data version of CS-2 used? If the authors use baseline-C/D data, the thresholds for lead detection from Tilling et al. (2018) would be used. Also see in Xiao F, et al. 2021. SCIENCE CHINA Earth Sciences.
5. The extent of snow depth and SIT distribution for W99 is smaller than others, please give some explanations.
6. For Bremen snow depth, the results for total sea ice and MYI in months from November to February are missing. Explanations also needed.
7. Table 1, W99 should be modified W99. Also for Table 2.
8. Figure 10, SIT results for the Canadian Arctic Archipelago are not shown.
9. L362-L364: The amount of OIB data involved in the training should be given here.
10. In reference 35, there is good coherence between the snow thickness data generated by the snow depth model trained by the neural network and the OIB snow depth measurements (correlation >0.8), whereas in this paper the correlation coefficient is only 0.55, what could be the reason for this?
11. L412: The effect of the radar penetration of CryoSat-2 on the snow is considered here. Please briefly explain how the factor of 0.22 in Equation 4 is derived.
12. L565: Why the correlation coefficients between sea ice thickness and OIB sea ice thickness retrieved from different snow depths are low (the highest correlation is 0.43)?

Author Response

The authors would like to thank the editor and the reviewers for your comments to help to improve the manuscript. We have now carefully reviewed and addressed all of comments which we hope meet with approval, with revisions to the manuscript shown in red. Please check word file.

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Review of

An improved retrieval algorithm of Arctic snow depths with AMSR-2 and its application on sea ice thicknesses of CryoSat-2 data

by

Dong, Z., et al.

Summary:
This paper introduces an alternative approach to estimate snow depth on Arctic sea ice from space-borne microwave radiometer data of the AMSR2 sensor. The approach introduced is a long short-term memory (LSTM) deep learning algorithm which uses 4 AMSR2 channels as input and which is trained with Operation Ice Bridge (OIB) snow depth estimates of its snow radar of years 2013-2018. The paper presents the results obtained with the LSTM method in comparison to five other snow depth products, two of which are modified climatologies, three of which provide actual snow depth information also based on space-borne microwave radiometry and one of which being based on differentiating elevation measurements of a laser and a radar altimeter. The results are also compared against OIB data of April 2019 - again together with the 5 other products. In addition, the paper provides an intercomparison of the effect the six different snow depth products have on radar altimeter based (Cryosat-2) sea-ice thickness retrieval. This is done via a qualitative intercomparison of the resulting sea-ice thickness maps and via a more quantitative intercomparison with OIB-based sea-ice thickness estimates.

The paper is in general well written and provides some interesting information in the part where it inter-compares the different snow depth products.
However, the paper provides an insufficient description of the LSTM itself and of how one ends up at the results obtained and presented in this paper. The paper lacks an evaluation of the LSTM itself. While it is stated somewhere that 80% of the input data are used for training and 20% are used for the evaluation of the method, the paper does not show any results; neither are the links to true physics compellingly laid out nor are issues such as overfitting discussed.
The paper provides an interesting discussion of the LSTM results but the outcome of this discussion neither justifies the title "improved retrieval algorithm" nor does it justify the inclusion of the section discussion the impact on the sea-ice thickness retrieval.
Neither did I find an illustration where, e.g., LSTM snow depth values are shown along an OIB overpass to enhance the credibility or where for certain locations in the Arctic ocean time series of LSTM snow depths are presented and discussed with respect to their consistency, nor did I find a critical discussion of the OIB data, their accuracy and validity and the potential impact of (always) taking these as the truth for algorithm developments such as the one carried out here.
In my view the authors tried to include five steps into one. Development of a new algorithm requires a careful evaluation of the steps and a careful and (self-) critical evaluation of the results against truly (!) independent data before examples of applications can be given.

General comments:
GC0 (purely editoral): Try to write as much as possible in active voice: We did this ... we found this ... we used ... instead of passive voice.

GC1: I recommend to change the title towards getting away from "improved" and getting away from presenting CryoSat-2 data. By the same token I recommend to keep the entire part of the CryoSat-2 sea-ice thickness retrieval and intercomparison for a second paper and delete it here.

GC2: The description of the data sets used needs to be improved. Please my respective specific comments.

GC3: The description of the LSTM method is not sufficiently clear and appears not to be complete either. Please see my specific comments. In addition: I don't find any evaluation of the method itself. 

GC4: The description of the co-location of the various data sets is not clear enough. Also it is not clear whether for the intercomparison (Table 1, Figure 6) and also Figure 7 all results are based on the area all products have in common. I note from Figure 5 that the Kwok product potentially has the smallest coverage and it is not clearly stated in the paper that this is in fact the regions where all the products were compared with each other - which should be the case.

GC5: In section 4 you immediately "jump" into showing snow depth maps in comparison to other products. This is not a credible way when introducing a new algorithm. I gave examples of what kind of additional illustrations of the LSTM snow depth you could use to make it credible in my text leading to the general comments.

GC6: The figures showing the snow depth results can and should be discussed in-depth. There is a lot that can be said and discused in light of existing literature. As pointed above: A critical discussion of the results is missing. Guiding questions could be:
- How independent is my intercomparison with respect to the OIB training?
- How independent is my intercomparison against other satellite products that also used OIB training?
- What are the limitations of OIB snow depth data (MYI, roughness, thin snow underrepresentation, preferred MYI sampling)?
- What is important for a snow depth data product user? Is it high accuracy in a month where I have other information (OIB) or is it high accuracy where snow depth information is particularly scarce (Nov/Dec)?

Specific comments:
Lines 17/18: These snow depth products need to be explained better already in the abstract

Line 21: If these are independent from the training data you should express this here.

Line 79-82: I don't get what the differences between algorithms listed are. What distinguishes a "retrieval algorithm" from those using an empirical relationship? Actually, you refer to [28] as retrieval in the next sentence but assigned it to "regression algorithms" here. This is confusing and calls for rephrasing.

Lines 151-154: The description of the AMSR-2 data needs to be improved. For your polar applications equatore overpass times might not be overly relevant. What is relevant would be whether you used data from both polarizations, the time period for which you used the data, the native and the grid resolution of the data. You writing also insufficiently explain whether the daily averaged grid product contains both, ascending and descending data or whether there are separate products.

Line 156: Projection at which latitude?

L158/159: Please be more specific here. What kind of a "spatial window" did you use? Where is its center located? Is the temporal window of 1 day simply a calender day or is it one day-period centered at a (which) time? The collocation of the data needs to be described in more detail.

Lines 160-172: If you decide to keep the sea-ice thickness part of the paper you need to condense this paragraph and provide the information that is relevant for the paper - which is basically which kind of data you used from which source for which time period and which region. Also, you might want to inform the reader what kind of data the L1B data actually are. If you carried out co-location with other data sets then you could also place that information here.

Lines 174-183: 
- Please revise these lines in the context of what is relevant for your paper. Are the DMS data relevant? What are "interglacial waterways"? Are you sure the range resolution of the snow radar is 0.06 m? Please provide references for the technical specifications given for ATM and snow radar. Please provide information about the accuracy (which is different from the technically defined vertical resolution). Please provide information whether the spatial resolution given for ATM and Snow radar is the same along- and across flight-track. 
- "ICESat-1" --> "ICESat"
- You might not want to provide information about the Antarctic; note that OIB flights took place there in October/November so that what you wrote is not correct.
- A citation and/or URL where you got the data from is missing.
- What is the reason for using OIB data from 2013 onwards instead of using data from 2009 onwards? One could guess that this is related to the AMSR2 launch in 2012 but it would be helpful to write that down.
- Please motivate your choice of the year 2019 for the evaluation. Isn't using one particular year for the evaluation kind of dangerous with respect to the specific conditions that might have occurred in that particular year only? Wouldn't it have been more valuable to take a certain amount of OIB data data choosen arbitraily from OIB overpasses of all years for the evaluation?

Lines 183-189: 
- "adopted" --> "used"; "to establish" --> "to develop"; please make clear that the reader understands that you are talking about your new snow depth retrieval method, hence be more specific.
- "sea ice freeboard" is not "measured" by OIB. It is a quantity derived from total (sea ice plus snow) freeboard measured by the ATM and the snow depth derived by the snow radar. Please correct your writing accordingly.

Lines 189/190: Please add information about the time period of data we look at and whether snow depths of intersecting OIB overpasses are averaged or simply overplotted with the most recent data set.

Figure 1:
- Please enlarge the figure
- Please use a color table that does not include the background color
- Please avoid using a color table that is often used for anomalies.
Please see also my comment to the respective text with respect to whether most recent OIB snow depths overplot old values or whether averaging is applied.

Lines 194-202: It is not sufficiently clear whether you worked with the original Warren et al. (1999) climatology or with a modified version of it. Please be more specific (and careful) in your writing. When you modified it by using a factor of 0.5 over FYI then please call it "modified W99" or similar. I recommend to not use the word "corrected" 
- Please provide information about the spatial resolution that is possible with this climatology; which did you use?

Lines 206-210:
- Please describe in more detail what that "regional weight factor" is. Isn't this quite similar to a FYI / MYI discrimination? Which ice type data set did they use (if they used one)?
- Is the AMSR-2 snow depth product used here the same as the one in 2.4.3? If not what is the difference?
- How are climatological data (W99) combined with the actual observations? Doesn't this mean that FYI areas have a true inter-annual variation while all MYI areas don't?
- I don't understand the last sentence of this paragraph.
- What are the spatial and temporal resolutions of the AWI snow depth data set? What is the grid? Where can I download it?

Lines 212-216:
- I recommend to make clear that the novelty introduced by Rostosky et al. was that they tried to expand the snow depth retrieval onto MYI - an issue which wasn't possible before. The basic concept of their algorithm is the same as that of the original algorithm by Markus and Cavalieri (1998).
- Instead of "vertical channels" write "vertically polarized channels" of "channels at vertical polarization".

Equation 1 and its entire explanation is a degree of detail possibly not needed here in the context of introducing the Bremen product. If you used that in your new approach then you should mention this issue there - with reference to the original Markus and Cavalieri (1998) publication.

Lines 226-228: How many OIB data are actually from March, how many from April? Isn't the fraction of March OIB data rather small and hence the validity of the Bremen snow depth for March limited? 
- Please also provide an estimate of the uncertainty of their product.
- What is the spatial and temporal resolution of this product? What is the grid? How did you colocate it with the other data?
- Please check whether "voyage" could perhaps be replaced by "flight"?

Lines 229-233:
- Compared to the other data products this is a very short description for a super innovative emerging state-of-the-art product. It deserves at least 1-2 more sentence describing the concept.
- What is the temporal, spatial resolution and the grid of this data? Where can I get it from?

Lines 234-243:
- This paragraph is unspecific in terms of which data (?) will actually be compared? It reads as if you only compare a model with other algorithms ... but how this is done is not explained. Please provide more information.
- In addition the degree of details provided (Keras ... TensorFlow ... version) is possibly not relevant for you paper and could be omitted.

Lines 244-252: Please provide more information and more motivation for using the data introduced here. In particular:
- What kind of SIC product did you use from EUMETSAT / OSI-SAF. Several are available; please provide the respective URL / information about the respective product. What is the spatiotempiral resolution and grid used?
- What kind of sea-ice type product did you use? On which data is it based, where can I get it, what is its spatiotemporal resolution and grid?
- You introducted Cryosat-2 data before. Why do you repeat this here?
- Where does the original snow density data modified by Mallet et al. come from? Please provide information about why this snow density correction should be used also in your paper. Please mention the number of the equation of [47] used to end up at the densities shown in Fig. 2.
- What do you need the DTU18 MSS data for? Not clear. Where exactly can I get the data?

Figure 2: Either here or in the text you need to make clear better how you derived these numbers, why you derived them and why you derived them for the period shown.

L258: I suggest to mention that it is typically not the pure brightness temperatures that are used here but brightness temperature ratios such as the gradient ratio.

Lines269-271: This is the place where that respective equation and its explanation should go.

Lines 282-287: These lines can be shortened byy simply mentioning that all T_b values are corrected with respect to the influence of water using equation 1.

Line 301:
Usually the letter sigma is used for the standard deviation. Is this what is meant here? If not I suggest to use another letter.
"tanh" ... stands for the  tangens hyperbolicus function? If this is the case then you should write it without a " " in the text ... and perhaps motivate its usage.

Lines 301-303: What do f, i and O describe mathematically? Are these vectors? Matrices? What is their dimension? Same for W and b ... what is their dimension and how are these computed?
What does the subscript "t" stand for? Is it time? Or simply the number of iteration steps?
Does the dot in the equations above denote the matrix product?

Line 305-307: Not clear what the loss function is. How is that mean absolute percentage error derived?

Line 307: Do you mean the "number of epochs is 250"?

Line 307: Please provide a reference for the Adam algorithm.

Line 310: "adopt" --> better "keep"

Figure 3 / description of the LSTM: I would appreciate if you could be more detailed and exhaustive in the description of the LSTM. I have difficulties to understand where the OIB data and where the TBs enter the LSTM. I have difficulties to understand why there are outputs h (is this the resulting snow depth?) at different steps of the processing chain and what these tell us. Which h-value would I take? Computation of W and b is not clearly enough described.

Section 3.2: Note: All comments to this section can be ignored if you (hopefully) decide to not show sea-ice thickness results and intercomparisons in your paper.
Please clearly write which parts of the retrieval of the SIT described here are new and developed by you. Parts / settings that have been developed by others should be marked clearly. It might make sense to condense this section towards saying that the SIT retrieval follows the approach of XXX with the following modifications made by you (including the motivation why you made these modifications).

Lines 319/320: Limiting the retrieval to SIC larger than 70% is NOT motivated by the attempt to discriminate between floating ice and leads. Present-day SIC products are unable to discriminate between leads and closed ice. Please take a look at the literature to find the real reason. One likely reason is that freeboard retrievals start to become unreliable in the presence of too much open water exhibiting a similar signal as the sea ice when rough. Another reason could be that ocean swell traveling into the ice cover causes surface height variations that can cause artificial variations in freeboard height.

Lines 335-339:
This reads a bit complicated. It might be easier if you introduce that you require the leads to determined (or approximate) the local sea surface height which is in turn required to compute the sea-ice freeboard. I am not sure using terms as SSHA aids in the understanding. Again - I assume that these are parts that have been described elsewhere and do not have to be repeated here.

Equation 13: Is there a reason why you denote the multiplication symbol with a star in equations 11 and 12 but with a cross in equation 13?

Please refer to Fig. 2 to make clear that you are using that snow density - which I hope you did.

Lines 366/367: Please rephrase the geographic description as this sentence reads as if the snow depth is retrieved on land.

Line 378: This is not correct. See the Kwok et al. paper in J. Geophys. Res. from 2020 on this topic, their equation (4).

Lines 381-383: Please re-phrase. It is not clear what you want to state here. Please take into account three things. 1) Snow depth products heavily relying on brightness temperature combinations that are sensitive to the presence of MYI (i.e. 19 and 37 GHz) might show deep snow in the presence of MYI even though no deep snow exists. 2) Snow depth products using brightness temperature (combinations) sensitive to surface roughness as well might show thick snow in regions where ice is actually deformed but snow depth not yet thick. 3) Saturation effects apply for deep (about 40-50 cm) snow packs - known for products using 19 and 37 GHz.
Note that it is very likely that snow depth products relying on methods such as deep learning or neural networks might appear to show snow depths thicker than the saturation snow depth but it is relatively clear that these products show these for the wrong reason as the physics behind the measurements simply does not allow to obtain signals from deeper within the snow than the saturation snow depth.

Figure 5: Very interesting!
- AWI and modified W99 are very similar over FYI - suggests that the PMW snow depth climatology used agrees very well with 0.5 * W99 snow depth (evidence from literature?)
- All non-climatological products but Kwok show a more or less clear distinction between FYI and MYI areas. Why is this? 
- NN and LSTM agree in showing far too large snow depths in Nov/Dec over MYI. This applies in particular to LSTM.
- NN and LSTM agree in showing little seasonal development of the snow depth over MYI but also over FYI; it is not clear why NN snow depth decrease from NOV/DeC to FEB over most of the FYI.
- KWOK has lowest snow depth overall over MYI.

Lines 405-418: I suggest to let Figure 6 speak for itself and only note that Table 1 provides the summary of the results shown therein.

Figure 6:
- modified W99 and AWI should be the same over MYI. Why is this not the case?
- Can you confirm that the bars shown in this figure are computed using the only the area which all products have in common - which actually is the one by Kwok? I actually doubt this is the case because AWI overall is considerably smaller than modified W99.
- Best agreement between all products is in April - the month with the best coverage with OIB data used for training, evaluation and consistency check of  4 of the 6 products.
- Kwok provides the most realistic seasonal increase in FYI snow depth and agrees well with the PMW products in march/april for FYI. All PMW products underestimate the seasonal cycle of FYI snow depth.
- How come LSTM snow depth decreases overall from Nov through Feb. while it increases for FYI and MYI? How come LSTM snow depth in March equals Kwok when being smaller by 2 cm for FYI and larger by 9-10 cm for MYI?

Lines 428-430: "stratified" makes sense once you have swapped x- and y-axis. The explanation "mainly affected by the correction coefficient" I don't understand. Please try to be more specific what you mean or avoid attempts to explain this observation.

Line 432-434: See my earlier comment on this statement which is wrong.

Line 437-442: This statement is hypothetical and is not backed up by your results. You could strengthen it by training the LSTM with the same data used for the Bremen and or the NN snow depth product to see whether your hypothesis is correct.

Lines 443-444: The fact that the LSTM retrieval is credible from the perspective of the retrieval accuracy when you only compare its results with OIB data from April 2019 when it has been trained with OIB data from mainly April 2013-2018 is not surprizing and only an indication that the training was successful. Looking at the other months clearly shows that LSTM is far less credible than you attempt to state here. Actually, for model development it is far more important to get the snow depths at the beginning of winter correct to allow for reasonable sea-ice thickness growth. Here the LSTM retrieval does not provide useful information.

Table 1: 
- Organization of the rows is a matter of taste but I would find the table more readable if it would provide the order ALL, FYI, MYI with the date (first column) shown together with ALL.
- I suggest to expand the caption and write more details, for instance:
"Overview of the monthly mean snow depth values of all six products for winter 2018/19. For every month, we show the mean and the standard deviation in meters separately for the entire (*) Arctic (ALL), first-year ice areas (FYI) and multiyear ice areas (MYI). No values for ALL and MYI are available for snow depth product "Bremen" for reasons detailed in the text."
You need to specify in addition (and this is because I added the (*)) what you mean by "entire Arctic" because this is only a subset of it determined as the area all products have in common (possibly the Kwok one).
- Finally, I recommend to use the same number of decimals AND to pay attention to the accuracy of the snow depth products. None of these provides snow depths as accurate as a millimeter. Therefore, also in the mean values and their standard deviations, using the millimeter decimal suggests a degree of accuracy that is not supported by the data. The first three FYI entries would hence read 0.10 +/- 0.02 for W99, 0.09 +/- 0.04 for AWI, and 0.14 +/- 0.04 for Bremen.

Figure 7:
- I would not call this "Validation diagram"; these are scatterplots.
- As the snow depth products are the dependent variables and the OIB snow depth represents "the truth" you should switch x- and y-axis and show the dependent variable (the products) at the y-axis as a function of the "truth" (the OIB data) at the x-axis.
- Please indicate the total number of data points. Is it the same for all six products?
- While all remote sensing products provide the snow depth at the grid resolution used (more or less), the OIB data points shown here represent an average over an unknown number of individual measurements. Would you mind providing this information as well? I note in this context, that little information is provided in terms of the spread (standard deviation) of the individual products.
- I suggest to make the dashed diagonal less thick and write in the caption that this is the 1-to-1 line.

 

Typos / editoral comments:
Line 38: "albedo effect" --> possibly you refer to the ice-albedo feedback? Write it like this please.

Line 40: What is "transmission of salinity and fresh water"? Not clear, rephrase.

Line 42: "As snow develops and accumulates" --> A snow COVER develops but snow accumulates ... unless you refer to all kinds of snow metamorphism (which I doubt was your intention). --> Rephrase.

Line 44: "weaken" --> "reduce" or "decrease"

Line 46: "As snow melts, it forms melt ponds" --> Well, not necessarily. Melt ponds may form. Also, there is snow melt on Antarctic sea ice which does not result in melt pond formation. Hence: Rephrase please.

Line 49: "to understand scientific issues related to climate change" --> please rephrase. It is not clear which specific role of the snow depth on Arctic sea ice you are targeting here. What are the scientific issues?

Line 53-54: "Satellite altimeters ... SIT" --> really? What is the accuracy? Can we obtain daily maps of the sea-ice thickness covering the entire polar region? --> I suggest to rephrase along the lines that satellite altimeters have proven to provide information about the sea-ice thickness distribution in the Arctic [references] and the Antarctic [references] - hence being more specific and less global.

Line 57: "sea ice freeboards" --> You use "sea ice freeboards" here as if this is the only parameter obtained. It is correct that a radar altimeter aims to measure the sea-ice freeboard but often fails to do so. A laser altimeter, in contrast, measures the total (sea ice plus snow) freeboard with higher accuracy. Therefore: rephrase please.

Line 64: What is "sea ice range" ? And what is meant by "on MYI" in this context? Not clear.

Line 82: I suggest to put the number referring to the citation directly behind the authors in the text. This applies to Markus and Cavalieri and many other occasions where you refer to the work of others.

Line 88: "is used"? --> perhaps better "is distributed"?

Line 89: "releases" --> "released"

Line 90-93: Is this lack of data over MYI limited to AMSR-E / AMSR2 data? Or does it apply to the DMSP-family of sensors as well? Rephrase if need be.

Line 149: "The" --> "the"

References: 
- Some of the references with more than two authors are listed with the full list of authors while others are not. This should be harmonized towards the rules applied by MDPI Remote Sensing.
- For some publications the year is missing.

Reviewer 2 Report

An LSTM based algorithm was proposed in this study to estimate the snow depth over the Arctic sea ice based on the Operation IceBridge (OIB) snow depth data and brightness temperatures data of AMSR-2 passive microwave radiometers. The algorithm is promising, and the paper is well organized. There are some issues that need to be clarified before its publication.

 

1, Six products were compared with OIB data including the snow thickness product generated by the LSTM method. Are the spatial-temporal resolutions of these products the same? If not, how to unify the temporal and spatial resolution between different data sources when verifying different snow thickness and sea ice thickness products by OIB airborne data?

 

2. The verification based on OIB snow thickness observations in April 2019 shows that the accuracy of the LSTM based snow thickness retrieval algorithm is higher than other algorithms. In the retrieval process, the authors estimated the sea ice freeboard by using Cryosat-2 data. If the sea ice freeboard is replaced by Cryosat-2 product released by AWI, will the LSTM algorithm still have the highest accuracy in sea ice thickness and snow thickness estimation? Can the authors carry out some comparative experiments to illustrate this question?

 

3. Due to the lack of in situ measurements, the validation of the proposed algorithm was carried out by comparing the retrieved snow thickness and sea ice thickness with the OIB data in April 2019. The retrieval results from November 2018 to March 2019 were not verified. The spatial distribution and temporal change of snow thickness and sea ice thickness should be investigated.