Review of the submitted manuscript; “Mapping the extent of giant Antarctic icebergs with Deep Learning”. The manuscript investigates mapping giant icebergs around the Antarctic continent using the deep learning convolutional neural network architecture U-net for semantic segmentation and compares it with other approaches, namely the Otsu thresholding and K-means segmentation.

Thank you for a well-written manuscript with strong English and interesting results covering an interesting segmentation topic. To summarise, there are a few things that need clarification, particularly in relation to the data used for the training. Otherwise, the manuscript is of high quality.

Major Comments

L86: Where do you have the resolution numbers from? Are you referring to SAR resolution or SAR pixel spacing because these are different and not identical to optical imagery? You should also mention the resolution type (high/medium etc..) Here is an overview of the Sentinel-1 products with resolution and pixel spacing. https://sentinels.copernicus.eu/web/sentinel/technical-guides/sentinel-1-sar/products-algorithms/level-1-algorithms/products 

L89: I think you should include more details on these preprocessing steps, e.g. I suspect you use the IPF (Instrument Processing Facility) denoising technique but it would be useful to know the version. I don’t know much about radiometric calibration but a couple of words about it could be useful. Is the multilook 6x6 or 1x6? There is also already some multilooking applied if you are using GRD. Have you considered this?

Fig. 1: missing legend on colorbar (year). In addition, the black dots/squares are a bit difficult to see. I think you should consider adding either a larger indicator or a square with the four corner coordinates of your images. Particularly for the Eastern red line, it is difficult to see the indicators.

L204-209: reading this makes me confused about how your training, validation and test setup are divided. It may help to elaborate on what type of cross-validation you apply. As I understand it, you select a test set, which is one of the icebergs and all the images associated with it. But this changes somehow (at least the number of images varies). Between different models or epochs? In my view, the models should be tested on the same data? 

I also think it would be useful to present some information about the data distribution of icebergs, e.g. how many images of icebergs are in each iceberg group, i.e. dark icebergs, coast etc. It would also be useful for the reader to understand how many images are available for each iceberg (so they are ready to examine whether there are any biases in the data).

Revisiting this comment, I realise that Table 2 has information about the number of images, I think it should be elaborated that the information is available here. I also see that Table 3 has similar information about the groupings

Minor Comments

L59-L70: This paragraph contains information about why you choose to utilise the U-net architecture. First, the benefits of it are highlighted followed by explaining that the paper Baumhoer et al. (2019) used to delineate ice shelf fronts with good success. The following sentence uses this as an argument for choosing the U-net. This makes it sound like it was only because of Baumhoer et al. (2019) that the architecture was chosen but I suspect this is not fully the case. I think I would rephrase it to something like “because of the many successful studies using the U-Net including one on a very similar topic (Baumhoer et al. (2019)), we choose to use the U-Net.

L63: For the authors information, there is a new entry of Stokholm et al. 2022 available, which also applies the U-net architecture. https://www.nature.com/articles/s41598-023-32467-x 

L71: I think you should consider splitting this section into a separate data section and a method section.

L82: which is frequent over the Southern Ocean? If it is clouds then it needs a reference. Alternatively, you could mention that clouds have a very similar albedo to ice (bergs) making them difficult to discriminate between in optical imagery.

L90: I would like some more details on why only the HH channel was used. HV for instance is much better at differentiating between ice and water and is less affected by the measurement incidence angle. (I am not gonna ask you to redo your experiments with both channels).

L99: I think you should reformulate the sentence to not include “cannot only”.

L101: I think you should explain briefly why icebergs are given certain names (e.g. B or C and number).

L103: what do you mean by “normal resolution”?

L121: You should add that the signal is depending on the surface properties, such as roughness, dielectric properties etc. Typically the angle of the satellite overpass is called the incidence angle of the measurement (also relevant in L130).

L140: as I read this sentence, it seems like images that have coast are grouped into the dark icebergs category but there seems to be a group called coast. It is a bit confusing to read.

L137-L153: this section was difficult to read. Lots of complicated sentence structures.

Fig 3. Just for your consideration, this is a very large figure, I think it could be smaller e.g. by yellow/grey squares smaller. I would only consider swapping the red and green arrows as they are most often represented that way for the U-Net, e.g. Ronneberger et al., 2015 original paper.

L171: Is the connected component analysis applied as part of the training, i.e. when calculating the loss and performing backpropagation? Or just as part of the results?

L220-222: it says initially that the supplementary material shows the best visualisations of the results but the final bit of the sentence says all outlines for the 191 images. I guess it includes the outlines.

L235: It would be useful to report the accuracy for icebergs alone, i.e. True Positives / Total Positives. Though it will have some uncertainty associated with it as the manual delineations are not pixel-perfect.

L241: seeing as I am a bit sceptical of the numbers in your pixel spacing, I would like to know what you define as the “pixel area”.

Table 1. I think the boldness of the best results should be further highlighted (thicker bold) if possible.

L293-295: I also suspect that increasing the receptive field of the network could have the model in these cases where the iceberg is relatively large.

Tab. 2, again I think the bold font could be increased in thickness.

L444: I agree, a larger dataset with many more examples of icebergs would be very useful to advance the training. Considering larger receptive fields may also help. Using data augmentation could also significantly help to create more varied training data for the models.

L462: the link does not work.

L494: the link does not work (https://doi.org/ appears twice)

L510: the link does not work

L515: the link does not work (https://doi.org/ appears twice)

L545: the link does not appear to be a hyperlink

In addition, there is a pdf document attached with grammatical suggestions.

The manuscript presented by A Braakmann-Folgmann and co-authors put forward a U-net approach to identify giant icebergs on different Sentinel-1 images which contain a variety of different environmental conditions. The resulting output from the U-net is then compared to manual delineations and other thresholding-based techniques (Otsu and k-means). I have made comments to try and not cross over with those made by reviewer 1.

The manuscript is generally well-written and structured, with only some tweaks required. The U-net performs well in some scenarios and provides scope for potential ‘giant’ iceberg identification on SAR imagery if training data is suitable. I very much enjoyed reading this manuscript and I am always excited in seeing new approaches to identifying icebergs on satellite imagery!

While this manuscript is relatively close to publication and mostly requiring minor comments (below major), I have some key comments, the major ones stated below:

• I’m not convinced such a complex approach such as a U-net is necessarily required to identify the largest iceberg in SAR imagery as it fails (using F1 metric) in two of the six environments (dark and coast). The U-net also has lower F1 scores in two of the six environments (fragments and open ocean) when compared to Otsu and k-means. This means out of six image classifications, only two environments (sea ice and other bergs (notably)) require a U-net approach which yield the highest F1 score. In my opinion, this suggests that the U-net devised is fit for purpose in only certain SAR images and is not necessarily suitable in others (maybe due to the amount of training each environment had?), as well as other approaches yielding better F1 scores. I appreciate in other metrics (false alarms particularly) the U-net does a better job than the other approaches. In only one environmental condition in Table 3 (‘other bergs’), the U-net outperforms the other approaches across the board (i.e. higher F1 score, with lower misses, false alarms, MAD). I think the narrative of the manuscript should consider that while the U-net can work suitably in some conditions, considering the amount of training it requires (and inability to identify larger icebergs than the ones identified in training) there should be caution made with regards to the success of this algorithm. I do however appreciate that the authors have been sensible with their claims and that the U-net requires more training data to be applied elsewhere.
• I’m hesitant to suggest that this approach is ‘automatic’. While the segmentation is automated in the images, it requires training data which is still reliant on manual input, as well as downloading/pre-processing the SAR data used in the manuscript.
• Two paragraphs (one in introduction: L44 to L58, one in methods: L121 to L132) could be merged to help re-structure one of the introductory paragraphs and remove one of the methods paragraphs which is introductory information.
  • The paragraph contained in the introduction (L44 to L58) would really benefit with some re-structuring (see minor comments below to try and help) to best describe the challenges faced by automatically identifying icebergs. It is currently difficult to follow with study examples.
  • The methods paragraph (L121 to L132) is a really nice piece of text explaining the difficulty of identifying icebergs on SAR images and would complement the introductory context. In its current state, this paragraph does not describe the method at all, or even the grouping classification, as this happens in L137 to L153.
• On a similar note, the first paragraph in the results (L227 to L256) reads as methodological which describes the statistics and comparisons of the different approaches. I would suggest moving this paragraph into a methods sub-section and start the results section where currently L257 starts (‘Comparing the performance of all three techniques…..’).

Minor comments:

• Avoid using both ‘iceberg’ and ‘berg’ interchangeably. I would suggest sticking with ‘iceberg’ and change any existing ‘berg’
• L11: Could the size be put in brackets to classify what giant actually is, i.e. (> 20 km²) - or is this just a general term?
• L12: Replace to ‘second’
• L13: Put in brackets what the two approaches are for segmentation
• L26: Change ‘derive’ to ‘delineate’
• L27: Add the references to first sentence if you’re stating a number of methods have been used
• L27: It might be worth mentioning within the sentence the specific approaches used which will set the scene for the upcoming paragraph, i.e. ‘A number of methods have been proposed, including thresholding (ref), edge detection (ref) etc.
• L28: Remove ‘simple’
• L31: Does ‘approach’ need to be added after ‘based’?
• L33: Would not include ‘etc’ --- either tell the reader what the parameters are or discard
• L36: Would avoid terms like ‘sophisticated’, ‘simple’, ‘elaborate’ as all approaches have their uses and people have different opinions on what algorithms should be classed as
• L38: Add ‘similarly’ before ‘Collares et al’ to connect the two sentences
• L38: Is it worth clarifying what k-distributions (as mentioned on L31) and k-means are? It currently reads as it is assumed the reader has prior knowledge of these approaches
• L39 to L42: Koo et al. paper is a very nice example again following the previous ones.... I think it is worth noting when discussing this paper the 'target' iceberg they track (B43 in this case) is originally delineated by a manual operator, limiting the application in terms of automation to other icebergs. I am pretty sure that is the method applied by that paper (just check), but it might be worth mentioning as it further bolsters your argument of limited approaches for 'true' automation
• L42: Rephrase ‘elaborate’
• L43: What does hand-crafted features mean? Are these manually delineated iceberg outlines?
• L44 to L58: This paragraph is the one mentioned in general comments about merging with the methods paragraph. I would suggest starting the new paragraph with a clear sentence stating what the exact problems are with 1) automated iceberg detection algorithms and 2) the application to SAR imagery. The examples can then be followed in the paragraph (i.e. L52 currently talks about ‘dark icebergs’ without explanation of what that term actually means --- this new re-structure will combine these approaches and hopefully clarify for the reader)
• L60: Would combine the first two sentences, so replace full stop on L61 with ‘which can outperform classic…’
• L68: Beginning of sentence could be rephrased to: ‘As the ice-ocean interface provides similar environmental conditions to an iceberg-ocean boundary…’
• L75: present in the image?
• L80: SAR needs abbreviating in the introduction (is L50 first use?)
• L91: The pre-processing aspect of the approach needs to include the sentence from L95 that it is conducted in SNAP
• L111: Put the temporal range of the dataset at the end of the sentence
• L112: Are the first 27 images rescaled as they are within the time period of B30?
• L116: Consider moving temporal range to L111 as noted
• L118 to L119: Final sentence is saying basically the same as the sentence starting on L116 – either consider merging or remove
• L121 to L132: Could be moved and merged with introductory information
• L134: Figure 2 could be placed under the new start of the section which actually describes the categories. Could scale bars also be placed on Fig 2 if possible, considering the size of these ‘giant’ icebergs
• L140 to L141: Rephrase to remove use of brackets from this sentence
• L146: Replace ‘bits and pieces’ with ‘fragments’
• L148: Rephrase sentence to try and avoid double ‘and’ if possible
• L155: The U-net requires the manual delineations, not ‘we’
• L156: Replace ‘click’ with ‘digitise’
• L156: How many iceberg outlines were manually delineated for training? This is a key part for training the network
• L165: Replace ‘click’ with digitise
• L173: Are there any statistics used to determine the Otsu threshold was the best suited, or was it visual quantification? Particularly important as the following sentence states Otsu has never been used for iceberg detection. Could this please be clarified
• L182: Apply or implement to be used instead of ‘suggest’?
• L193: Dash for ‘in-between’?
• L199: Rephrase ‘would like to discard’ to ‘require the removal of small icebergs…’
• L222 to L223: Struggle to follow this sentence, are you saying that your analysis combines both statistical and visual quantification to interpret the success of the approach? If so, could this please be rephrased
• L223: Consider rephrasing to ‘After an overall analysis, we assess the performance of the approaches for identifying each iceberg and…’
• L224: Different environmental conditions in the scenes? Rather than ‘challenging’?
• L227 to L256: Mentioned in general comments this is predominately methods which are important, but no results are provided
• L229: Add ‘an’ between ‘as iceberg’
• L227 to L256: I think it would benefit the study if it was clarified what an F1 score actually is and what it does (i.e. is 0 bad and 1 good? Is it a statistical comparison?) Would be easy to add a few sentences to describe the F1 score if the paragraph it resides in is moved to the methods above
• L239: Add references after ‘previous studies’
• L266: Considering the size of the icebergs which are being identified, it might be worth quoting the actual area difference between U-net and manual delineations, as well as the % difference as it will probably be a very large area difference (i.e. deviates by 12% which is X km²)
• L275: Does ‘dataset’ mean the number of available images for this iceberg? I would suggest clarifying this
• L277: Consider rephrasing to: ‘Furthermore, B41 remains in close proximity to its calving front for a significant period of time, which means…’
• L284: Casual phrasing ‘fine’, replace with U-net is ‘suitable’
• L292 to L293: Does this therefore suggest that U-nets are not always required to identify icebergs, rather only in images with certain environmental conditions?
• L305: Table 2 – Good to see how the U-net performs for the test dataset
• L310: Figure 4 – Really nice figure and shows the U-net varies in terms of success, depending on the image conditions
• L348: Is it not 'most' cases rather than ‘some’? While U-net does better than the other two approaches for coasts, it only has an F1 score of 0.34. I’d suggest this therefore means U-net struggles in most scenarios which contain termini?
• L361: I absolutely appreciate the fact land masks are temporally stagnant and therefore bergs in close proximity could be masked as well - however, it could be worth mentioning if the ice shelves frontal positions from Baumhoer et al. are available, they would provide a potential position to derive your own mask (for each scene and respective frontal position) which would be temporally dynamic and overcome this problem? A consideration which could be noted (I’m not saying do this by the way!)
• L372: Replace ‘not straightforward to compare’ with ‘not directly comparable’
• L408: Replace ‘humans’ with ‘manual operators’
• L413: See key comment about the algorithm being automated with the word ‘automatically’ used here
• L413 o L425: I think it is worth clarifying in the conclusion that the U-net is currently fit for purpose in certain image scenarios and not currently scalable to larger (and potentially smaller?) icebergs, however as noted with more training data, there is at least scope to assess the potential of applying a similar U-net to more SAR imagery
• L429: It might just be me, but the Zenodo link doesn’t seem to work (doesn’t seem hyperlinked)