Automatic Extraction of the Calving Front of Pine Island Glacier Based on Neural Network
Round 1
Reviewer 1 Report (New Reviewer)
Comments and Suggestions for AuthorsThe authors have used a U2-Net neural network model to automatically extract the location of the calving front of a few glaciers in Antarctica. A few major points of concern that I have after reviewing this paper include:
1. What is the dataset being used to determine if a U2-Net pixel result was a True Positive, TN, etc?
2. The number of erroneously classified pixels in the U2-Net result was described as "small", when it looks like it was actually a fairly large area in the ocean pixel block.
3. How was the manually interpreted calving front location derived? What is the minimum expected error associated with this? Both of these questions were hardly touched on.
4. The most important point of discussion that is missing and needs to be included is how this technique improves on previous efforts to map the calving front using neural networks. A quantitative comparison to other papers that have used optical and SAR imagery to map the calving front should be performed to highlight this technique's importance (and future adoption).
Please see the file attached for line-by-line comments.
Comments for author File: 
Comments.pdf
English language is mostly adequate. A few grammatical errors were noticed throughout.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report (New Reviewer)
Comments and Suggestions for AuthorsI have reviewed the manuscript titled "Automatic Extraction of the Calving Front of Pine Island Glacier Based on Neural Network". The paper compares three deep learning methods for extracting calving front locations using LANDSAT imagery. The overall content is comprehensive and sound. I recommend the paper can be accepted after addressing the issues I have raised below.
- The first use of the abbreviation 'FCN' in the abstract and main text should include the full terminology.
- Delete the last sentence of paragraph three in the introduction.
- Lines 171-173, explain the methodological details more thoroughly and provide references for this approach.
- Lines 180-181, provide citations.
I recommend it to invite a native English-speaking expert in the field for in-depth polishing.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report (New Reviewer)
Comments and Suggestions for AuthorsThis study presents a comparative analysis of the FCN, U-Net, and U2-Net models for the purpose of extracting the calving front of the Pine Island ice shelf, located in the southwestern region of Antarctica. This research contributes original and relevant insights to the field of remote sensing. It is worth noting that in the study area, the Filchner and Totten ice shelves were neither mentioned nor identified on the study area map. It would be beneficial to clarify whether these ice shelves are part of the Pine Island shelf and provide their exact locations.
Additionally, this paper underscores the effectiveness of the U2-Net method in classification tasks. While the methodology is well-structured, the abstract lacks information about the materials used. On line 123, it is important to clarify whether the high-resolution images from Landsat were used and provide the appropriate citation.
Furthermore, it would be beneficial to specify the software or codes employed in the application of the U2-Net and other models. The manuscript demonstrates a high level of scientific rigor, with an appropriately designed experiment that effectively tests the hypothesis. The results are reproducible and well-detailed in the methods section.
Regarding the figures, it is important to mention that figures 3, 4, and 5 were adapted from other publications and provide the appropriate citations. Finally, the conclusions drawn in the manuscript are consistent with the presented evidence and arguments, and the references provided appear appropriate. I will also check the additional references you mentioned:
"U2-Net: Going Deeper with Nested U-Structure for Salient Object Detection" by Zichen Zhang, Martin Jagersand, Chenyang Huang, Osmar R. Zaiane, Masood Dehghan, Xuebin Qin, published on May 18, 2020. (Please provide volume/issue or a URL if available).
"SKDCGN: Source-free Knowledge Distillation of Counterfactual Generative Networks using cGANs" by Sameer Ambekar, Ankit Ankit, Diego van der Mast, Mark Alence, Matteo Tafuro, Christos Athanasiadis. (Please provide the publication details or a URL if available)." “
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report (New Reviewer)
Comments and Suggestions for AuthorsThe manuscript is well-written. Some comments/suggestions for improvement are provided below.
General comments:
1. The manuscript uses an artificial Neural Network (ANN) to extract the Calving Front of Pine Island Glacier automatically. What is a novelty of your study comparing to other papers? Please write a few sentences describing the work's uniqueness.
2. In Line 161-163, the authors stated that "the images from 2014, 2016, and 2018 were subjected to edge filling and clipping, respectively, resulting in a final number of 1656 training samples." Is the data from 2014-2018 used for training here, and what about testing? More details are needed?
3. Please cite the source of all equations in the manuscript.
4. The calving front of the Pine Island ice shelf was calculated using three models: FCN, U-Net, and U2-Net. As shown in table 4, the U2-Net successfully found to be more suitable for remote sensing image interpretation in calving front extraction, outperforming the FCN and U-Net. The authors must verify/compare their findings with those of other papers.
5. Please include in the conclusion section the possibility of, what is the benefit of the obtained results, and in what fields can they be applied/utilized?
Specific comments:
1. Please use the proper reference style in Lines 71, 73, and 76. r
2. Figure 1: It is preferable to include a map of the southwestern region of Antarctica in the world map.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report (New Reviewer)
Comments and Suggestions for AuthorsThe authors sufficiently addressed my major and line-by-line comments in this revision. They addressed my major concerns about how the ground truth data used to assess classification accuracy was collected as well as how their results fit within the broader scope of delineating calving glacier fronts. They show that their technique improves on other very recent papers (2019-2022) by about 28%, representing a substantial improvement.
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
Comments and Suggestions for AuthorsSong et al. present a revised manuscript which had many flaws in the first version. Some changes were made which did not improve the manuscript. Main changes made for the revision include:
- Adding two authors
- Removing plagiarism suspicion.
- Adding additional information on the training of the neural networks.
- Two additional accuracy metrics (Accuracy, Cohen’s Kappa Coefficient) were introduced.
Comments to the answers of the authors referring to the numbers used in the response letter:
1. Ok. Nevertheless, I would highly recommend the journal editors to use a plagiarism check software to make sure the manuscript does not include any other indication of plagiarism.
2. Unfortunately, I cannot see an improvement of glaciological understanding. The statement in line 33 is wrong. Ice shelves are floating on the ocean. If this ice breaks up sea level won’t rise. It is the same as having an ice cube in a water glass. The ice cube melts but the water level remains the same. Sea level rises if active ice shelf area with buttressing forces breaks up. This will increase discharge from grounded ice and this kind of mass loss increases sea level rise.
4. For me it was not clear from the manuscript that the study area is solely Pine Island. To make this clear it would be wise to state this in the title: Automatic Extraction of the Calving Front of Pine Island Glacier Based on Neural Network
5. I will leave gramma and wording issues to the journal editors.
6. I welcome the more detailed information on the training process. That now shows why the training did not preform as good as in other studies (especially for the UNet). First, augmentation is missing. This is a crucial part to increase training samples and get a more robust network. Augmentation should be included to create reliable results. E.g. as already applied in the original paper by Ronneberger et al. 2015. Furthermore, the authors state they trained the network for 800 epochs. After so many epochs and so little training samples overfitting occurs. You need to monitor the training process with validation data and stop the training early enough before overfitting occurs. This will lead to different epochs for each network as you cannot use the same hyperparameters for different networks to identify the best performing model.
7. First, I would recommend to make this very narrow selection of images clear in the beginning of the manuscript. Furthermore, is your model is restricted to so many circumstances you should definitely increase your training and test set.
Further specific comments:
L91ff: This is completely off topic.
Table 2: Did you train for 800 iterations or epochs? That is not the same!
Comments on the Quality of English Language
I will leave gramma and wording issues to the journal editors.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe revised version of the article has fully addressed the issue I mentioned last time, and I believe it is acceptable for publication.
However, there are still some small issues that need to be addressed. Firstly, the positions of some images are not suitable and need to be adjusted to the appropriate positions. In addition, The revised version of the article has fully addressed the issue I mentioned last time, and I believe it is acceptable for publication.
However, there are still some small issues that need to be addressed. Firstly, the positions of some images are not suitable and need to be adjusted to the appropriate positions. Additionally, I suggest adding the word 'ice shelf' to the title, as this should be the Calving Front of ice shelf.
Comments on the Quality of English LanguageEnglish needs further refinement. I suggest finding a native English speaker to help improve the language.
Reviewer 3 Report
Comments and Suggestions for AuthorsI suggest, that you should focus on the title, which makes emphasis on methodologic proposition to detect ice shelf caving fronts based on neural network. It seems like you apply the suggested methodology to detect variations. At least this is what you present in your results. There is no evidence that your neural network approach is better than any other well established edge detection algorithm. The comparison which you present (FCN, U, U2), seems something artificial. You did not apply any type of ground truthing.
Furthermore, U2 generates artefacts (Fig 6) which by image processing approaches can be prevented.
Why did you propose the use of Landsat images? Why did you choose images observed during January, February and March?
Comments on the Quality of English LanguageYou should thoroughly review your document as it is written in poor English. Maybe somebody can help you checking grammar and style.
