The authors' effort in developing the method presented for fracture mapping with Sentinel images is highly commendable. If the produced dataset undergoes more careful validation steps as suggested below, it is expected to be a valuable asset for the cryosphere community. The potential application of the author’s result to enhance the representation of damage in ice flow models is important. However, I do think that a more comprehensive explanation and validation of the fracture detection methodology utilizing machine learning is necessary to verify the robustness of the method and its results. Below are three major comments:



Firstly, the evaluation of fracture detection presented in the paper is primarily qualitative. To provide a more thorough assessment of the algorithm's performance in detecting fractures, the authors should include quantitative measures. Although the authors mentioned sensitivity and specificity, they did not provide the actual values of these measures. To address this issue, the authors could obtain a labeled ground truth dataset through other automated methods, such as Izeboud & Lhermitte (2023), or manual annotation. The authors mentions manual labels “is likely to be uninformative given the subjective nature of producing manual annotations“ but all of the qualitative descriptions regarding “sensitivity” and “specificity” in the evaluation section essentially were based on the authors’ subjective judgment regarding what counts as surface crevasses, rifts, and basal crevasses. Therefore the authors already potentially impose subjective judgements. Therefore, it would be more transparent to provide annotated fractures that represent the authors' judgments, and calculate standard quantitative measures of neural network performance, such as sensitivity, specificity, area under the ROC curve, and F1-score. The authors can surely acknowledge the fracture annotation, just like any labels in the glaciology literature, could contain subjective bias. Assuming that fracture annotation is improved in the future, the future users can follow the author’s NN training/evaluation procedure to improve performance.



Secondly, to gain a better understanding of what this method captures in comparison to other existing methods, it is important to conduct some comparisons with existing fracture maps such as Izeboud & Lhermitte 2023 (also Sentinel image with a different method) or Lai et al 2020 (same method with lower resolution images). The code for Izeboud & Lhermitte's method and the fracture map produced by Lai are openly available. It is likely that this method complements existing techniques, as the authors' Unet captures the sharpest fractures, while Lai's map captures smoother/larger features visible in MOA images. The authors' method also appears to detect fine-scaled fractures, which are also captured by Izeboud & Lhermitte's method.



Lastly, the method used to generate training data is not well explained in the paper. It would be helpful if the authors could provide visual examples of the training data used in the 4-5 iterations described in lines 96-102. Additionally, the authors should clarify how a neural network used for detecting calving can eventually detect surface crevasses and even surface expression of basal crevasses that appear quite distinct from a calving front. The authors mentioned “manually selected images for which the network performed well at the task of crevasses detection to form an updated training 100 dataset“. However, they did not explain how they selected these images. Does this manual selection include some but not all basal crevasses-like features, so that the final NN represents basal crevasses with low but nonzero prediction probability? Again this training data generation step already involves subjective judgment that the neural network learns from. Therefore, it would be beneficial to provide visual examples of the training data to allow for a more thorough understanding of their methodology.



- Izeboud, Maaike, and Stef Lhermitte. "Damage detection on Antarctic ice shelves using the normalised radon transform." Remote Sensing of Environment 284 (2023): 113359.



- Lai, Ching-Yao, Jonathan Kingslake, Martin G. Wearing, Po-Hsuan Cameron Chen, Pierre Gentine, Harold Li, Julian J. Spergel, and J. Melchior van Wessem. "Vulnerability of Antarctica’s ice shelves to meltwater-driven fracture." Nature 584, no. 7822 (2020): 574-578.





Line by line comments:



Line 46: Re: “We produce continent-wide maps of fracture”. This appears contradictory with lines 136-138 where the authors mentions that large parts of Ronne and Ross ice shelves can’t be mapped.



Line 62: Please cite classical references, Irwin 2057, for the definition of Mode I, II, III fracture:

- Irwin, George R. "Analysis of stresses and strains near the end of a crack traversing a plate." (1957): 361-364.



Line 65: Add the following references that clearly demonstrate “basal crevasses which can result in visible large-scale depressions in the surface“:



- Luckman, A., D. Jansen, B. Kulessa, E. C. King, P. Sammonds, and D. I. Benn. "Basal crevasses in Larsen C Ice Shelf and implications for their global abundance." The Cryosphere 6, no. 1 (2012): 113-123.



- McGrath, Daniel, Konrad Steffen, Ted Scambos, Harihar Rajaram, Gino Casassa, and Jose Luis Rodriguez Lagos. "Basal crevasses and associated surface crevassing on the Larsen C ice shelf, Antarctica, and their role in ice-shelf instability." Annals of glaciology 53, no. 60 (2012): 10-18.



Line 66-69; line 84: Define “sharp/narrow.” The authors emphasize that sharper features are easier to detect with the method presented here, including surface crevasses, rifts and the “sharpest looking basal crevasses/narrow surface depressions”. It is unclear what are “sharpest looking basal crevasses”. Can the authors explicitly state what sharpness means, e.g. what are the characteristic sharpness for the fractures to be detected? Can the authors simply use a few altimetry elevation data to demonstrate sharpness over the rifts, surface crevasses and the “sharpest looking basal crevasses” identified using Sentinel 1? Wang et al. 2021's paper includes a few examples of the ICESat-2 data over fractures on Amery that shows the steepness of the crevasse walls.



- Wang, Shujie, Patrick Alexander, Qiusheng Wu, Marco Tedesco, and Song Shu. "Characterization of ice shelf fracture features using ICESat-2–A case study over the Amery Ice Shelf." Remote Sensing of Environment 255 (2021): 112266.



Figure 1a: How do the authors know 3 and 4 correspond to basal crevasses and surface crevasses based on Satellite image, given that they look quite similar on the imagery? The authors can check with ICESat-2, radar profile, higher resolution satellite image, or simply use basal crevasse locations that have been identified in the literature (e.g. Luckman et al 2012; McGrath et al 2012). Given that this paper is focused on fracture mapping and the word “basal crevasses” was mentioned several times, I believe it is important to justify the existence of basal crevasses in a few places where the author indicates the UNet identifies basal crevasses.



Line 101: Re: “1000 images”. Remind the reader how large each image is.



Line 110: Re: “a threshold can be applied to produce binary maps, with values varying from 0.3 to 0.5 depending on the features of interest.” Can the authors explain how these thresholds are determined?



Figure 3: Is “D” neural network output (mentioned in appendix) or fracture density?



Line 176-178: Re: “A formal validation of the accuracy of the crevasse mapping results presented here, for example, by comparing our maps with manually annotated satellite images, is challenging. This is especially true for our method which produces a continuous, rather than binary, output”  This is not entirely true. Classification with a probability output is a stanford ML task and there are several standard validation metrics used to evaluate the performance of models that produce non-binary output, such as the "Area under the ROC Curve.”

Line 190: Re: “the methods we have developed extract the vast majority of features in the backscatter images to produce our crevasse maps while highlighting very few features erroneously.” Without comparison between this method with ground truth, we don’t know if the method predicts very few features erroneously.



Line 222-227: The authors mention some smooth features, likely basal crevasses, are also detected with this method. As the method’s training data include only sharp edges, I’m puzzled how the NN eventually develops the ability to pick up smooth features. Was that related to the “bootstrapping” procedure to make training data?  



Line 258: Fracture density. For rifts, does the fracture map count the areas within the rift as fracture, or only the two “sharp-sides” edges? If the latter, the current definition of fracture density can underestimate the effect of fracture on reducing ice viscosity, as it doesn't include the void space in between the rift walls.



Figure 5: Should “Buttressing ratio” be “buttressing number”? Use kappa as defined in the appendix. Shouldn’t the buttresing number be <1? How can it be as large as 3??



Line 480: Clarify what resolution count as “the most-high resolution cases”



Line 490: Re: “There are, however, certain disadvantages to the use of SAR. Some ice shelf crevasses appear smoother on the surface than in optical imagery. For example, the crevasses on Dotson Ice Shelf, that appear faintly in SAR backscatter images at 50 m resolution, can be seen more clearly in the MODIS MOA image over the same region.” Are these fractures visible in MOA mapped in Lai et al 2020?



Line 500: Note that it would be extremely difficult to have a reliable crevasse-depth estimate as the crack tip can be generally narrower than the data resolution.

 

The manuscript presents a method for automatic detection of certain types of crevasses in Sentinel-1 imagery of the Antarctic Ice Sheet. Further, the derived dataset is introduced and first analyses are derived from the data. Of note is the monthly resolution of the data product and the near pan-Antarctic coverage. Overall, the manuscript is very well written sand shows clear promise. I believe that the resulting crevasse maps will be of high use for Antarctic research and allow for a better understanding of ice sheet dynamics. The main concern is with the description of the used methodology and the validation thereof. The scientific value of the derived crevasse maps, and in turn the analyses, hinges in large parts on the reliability of the used neural network approach, which should be more thoroughly evaluated.

Major comments:

1. Neural Network Training: The methodology used is not entirely clear. The bootstrapping approach is non-standard, so it might be helpful to explain it in a bit more detail. Currently, it is not entirely clear which criteria are evaluated to select the bootstrapping samples. Further, please clarify which calving front dataset was used for the initial training round.

2. Evaluation of the Model: Quantifying the accuracy of the proposed method is critical for estimating the credibility of the predicted crevasse maps and the subsequent analyses. While I agree that it is not possible to thoroughly compare it to human annotators on a pan-Antarctic+multi-year scale, it would still be interesting to see such a comparison for single scenes. Further, comparing it to simpler (non-DL) methods or existing DL approaches for crevasse detection (e.g. [1], [2]) might be helpful for the readers to better understand the advantages and drawbacks of the newly proposed method.  

3. Inspection of the crevasse maps uploaded as review assets suggests that the method is also sensitive to local changes in texture which are not related to crevasses, like the calving front or ice mélange. This should be discussed.

Minor comments:

* Lines 104f.: If my understanding is correct, two networks (N_A, N_B) are employed to map two types of crevasses (type-A and type-B), and then the results are stacked and a softmax is employed. Is this also done during training or only for inference? What is the benefit of employing two separate networks over a single, multi-class UNet?

* Line 106: The mention of "scalar outputs" is in contradiction with the softmax, which is a vector-valued function.

* Lines 131f.: The quality of the type-B detection depends on the availability of SAR acquisitions from multiple look angles.131f.). With the failure of Sentinel-1B in 2021, the number of available acquisitions has roughly halved. Does this affect the type-B crevasse maps?

* Line 177f: "This is especially true for our method which produces a continuous, rather than binary, output". This argument can be made for any neural network-based method, so it is not quite convincing as a reason for not providing quantitative evaluations. Further, metrics like AUROC exist for such cases.

* Line 181: Were the presented example SAR images hand-picked or randomly chosen?

* Line 528: "However, a greater number of crevasses can be seen at 10 m resolution". How readily can the proposed methodology be adapted to higher resolutions? Maybe a sentence could be added here outlining the ease/difficulty of adapting to higher resolutions.

Typos:

* Line 381: "Firstly, it likely"

[1] Lai, CY., Kingslake, J., Wearing, M.G. et al. Vulnerability of Antarctica’s ice shelves to meltwater-driven fracture. Nature 584, 574–578 (2020). https://doi.org/10.1038/s41586-020-2627-8

[2] Zhao J, Liang S, Li X, Duan Y, Liang L. Detection of Surface Crevasses over Antarctic Ice Shelves Using SAR Imagery and Deep Learning Method. Remote Sensing. 2022; 14(3):487. https://doi.org/10.3390/rs14030487