E-MPSPNet: Ice–Water SAR Scene Segmentation Based on Multi-Scale Semantic Features and Edge Supervision
Round 1
Reviewer 1 Report
Please see the attached pdf file.
Comments for author File: 
Comments.pdf
Author Response
Please see the attached pdf file.
Author Response File: Author Response.pdf
Reviewer 2 Report
This paper presents a new architecture, E-MPSPNet, with the aim of improving sea ice - open water classification. This is an important topic in improving the utilization of the large volumes of synthetic aperture radar (SAR) satellite data that are becoming available, and reducing the workload on the expert analysts at the various national ice centers. As the authors cover in their introduction, this topic has been covered by widely in the literature. They also note that previous initiatives in this field have suffered from the need for too much manual intervention and are affected by the quality of data, and are thus limited regionally and temporally.
The introduction could do with some further work, and there could be additional discussion on the quality and merits of the various input datasets for this task. It would be interesting to see this approach tried with alternative input datasets, ASIP in this case seems to be limiting factor in the examples shown. As a result there is not sufficient justification for the claim at L562 "our model is effective in edge detail detection", so I suggest some further work and major revision.
L36-37 There is a contradiction in the statement that melting sea ice results in a large number of floating ice or icebergs. This should be revised, I think what as meant here is that thinning of sea ice results in a more dynamic ice cover, and this can result in greater movement of the sea ice edge. In addition, the general retreat of the Arctic sea ice cover is exposing glacier fronts to open water conditions and waves, resulting in the calving of more icebergs.
L40 "research institutes". Replace "research" with "operational monitoring" as the examples listed are the national authorities for sea ice and iceberg monitoring for their respective Arctic regions.
L40 "U.S. Snow and Ice Center". In this context and operational monitoring this should be "U.S. National Ice Center". There is also the National Snow and Ice Data Center (NSIDC) but that is a data repository and not a sea ice monitoring organisation.
L44 "limited in accuracy and resolution". This is a controversial statement given the focus in the introduction on operational ice charting. It depends on the national center and its mandate for supporting users, also whether this is the routine ice chart production or studies for specific users. Do the authors feel that their proposed method adequately addresses this?
L63-64 "These features are extracted by manual interpretation". This statement seems out of context here. The classification of features are often from manual interpretation for training datasets but the derivation of the features themselves is typically automated.
L138 Replace "Sea Ice Operations" with "Greenland Ice Service".
L182-3 "800 x 800 pixels of patches" rephrase as "patches of 800 x 800 pixels"
L375, Fig.3 caption refer to "interglacial gaps". Interglacial is a geological term for the periods between glaciations. Replace with "open water gaps", or "inter-floe gaps".
Figs.3,4,5,6 a These look like single polarization images, not dual polarization.
L391-402 There is a lack of sea ice edge detail across all examples. Figs.4, 8c and 9c, would expect to see more fine detail in the ice edge as these are diffuse edges with patches disconnected from the main ice pack. It would also be expected to see open water areas within the main ice pack, and these are often important for some vessels that want to use these for navigation. Various studies, e.g. IICWG and KEPLER, have reported that maritime users desire sea ice information products with a spatial resolution of 300 meters or better.
The examples fail to show the algorithm performance in areas of coastal landfast ice that are prevalent around Greenland. These typically have low response in SAR images, and cause issues for other examples of ice-water classification.
L508, Fig.7 Please confirm that this is the "Baltic Sea", maybe show the position of the SAR image in it's wider geographical context. This does not look like a SAR image of Baltic Sea ice, which has distinctly different sea ice conditions, and it's not an area covered by the DMI ice charts.
L565-572 I agree, and this is an issue with the choice of ASIP as the input datasets. The DMI Greenland Ice Service has to cover a wide area and include sea ice type (stage of development) in their analysis, so only capture the broad details of the sea ice cover. Other national agencies focus on detailed mapping of the ice edge and their ice charts would be more appropriate for training an ice-water classification.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In this paper, the authors propose a new neural network-based model, called E-MPSPNet, for water ice segmentation in SAR images, which achieves superior results to other previously developed methods, with which this method is compared. The importance of each of the steps included in the model is also demonstrated by ablation.
The article is very well written, with only minor hyphenation errors and only occasional typos. Therefore, I consider that once the text has been revised, the paper can be published.
Two examples are:
Line 202: "predicted edge feature maps Finally": Missing dot.
Line 317 "learing".
Author Response
Dear Reviewer:
Re: Manuscript ID: remotesensing-1973446 and Title: E-MPSPNet: Ice-water SAR Scene Segmentation based on Multi-scale Semantic Features and Edge Supervision
We appreciate the reviewer’s positive evaluation of our work.
We have carefully considered all the comments and revised our manuscript accordingly. In the revised manuscript, the typos and grammar errors we found have been corrected, and the main revisions have been highlighted in red.
We hope our revised manuscript has well addressed your concerns and it can be accepted for publication.
Sincerely.
Reviewer 4 Report
In this study, the Authors design a new neural network to automatically segment dual-polarized C-band SAR images into open sea water/sea ice. Hence, this work falls within the research topic about the exploitation of satellite imagery to develop added-value products in the cryosphere domain. The proposed network includes the use of multi-scale semantic features and an edge supervision mechanism. An open data set, which was further augmented by rotation/flipping, is used for the experiments. Results are compared with several state-of-the-art segmentation methods based on neural networks and performance are assessed using a ground truth provided by sea ice charts from DMI.
Please find some main concerns and minor suggestions to improve the manuscript before considering it suitable for publication.
MAJOR COMMENTS:
1) The usage of English language must be improved thoroughly along the manuscript.
2) The original contribution and the novelties proposed in the network design with respect to existing literature must be clearly highlighted.
3) The description of the data set must be improved. Information as spatial resolution, incidence angle, area coverage, etc. must be included.
4) If I well understood, segmentation problem, i.e., a binary sea water/sea ice detection problem, is addressed in this study, see Table 1. However, results presented in Figures 3-9 show "classification results", i. e., four different color/classes (which are not specified) are present. Clarify this key point.
5) From experimental results, i. e., see Table 3, it seems that improvements with respect to the other methods considered for comparison fall in the range 1% - 2% depending on the metric. Does it really deserve to pay those small improvements in terms, maybe, of greater network complexity/computational burden? In simpler words, from the analysis and discussion must be clear the benefits obtained by the proposed approach and how they are "paid" (see also lines 413-420).
6) Figures 3 - 6 (a), do represent HH or HV SAR images? In addition, at a first look they seem to be speckled and, therefore, the interpretation of the open water/sea ice boundary is hampered. Did you consider the chance of pre-process SAR data using proper speckle filtering? Some of existing speckle filters were though to reduce speckle noise while preserving edges.
7) The analysis and discussion presented in subsection 5.2 and 5.3 is just qualitative and heuristic, there is no proof about this. The relevant conclusions are not supported by experiments that can be checked and, therefore, this part is very weak.
MINOR COMMENTS:
1) The title should not include the abbreviation proposed for the method.
2) Check the typo "[1]" just above the first word of the title.
3) Please define all the abbreviations only once, at their first instance (e. g., ASIP in the abstract, line 23).
4) Introduction just focus on Arctic, while in principle the proposed method can be generalized even in Antarctica. Talk about polar regions should be more appropriate.
5) Lines 50-54: I do not agree. Several studies agree in considering the cross-pol SAR channel sensitive to high wind regimes, see 10.1109/LGRS.2010.2085417, 10.1080/01431161.2015.1134845, 10.1109/TGRS.2014.2366433, 10.1109/TGRS.2013.2293143.
6) Line 57: Reference [8] does not deal with SAR imagery. Please check.
7) Line 144: what do you mean with "stripe noise"?
8) In point 3 of the methodology, see line 178, the Sobel operator is claimed to be used for detection. However, some studies as 10.5721/EuJRS20164912 and 10.1109/JSTARS.2020.3036458 found the Canny edge detector to be an "optimum" choice. Please discuss about this.
9) Line 185: please detail the meaning of "normalization" procedure and how it was applied.
10) Please add the total final number of SAR scenes/pixels used for experiments.
11) Subsection 4.1: how to set those parameters should be motivated.
12) It is not clear, from my side, if the Authors used as SAR image inputs both HH and HV channels or they only used the cross-pol one. Please clarify.
13) Line 497: what do you mean with "scene graph"? I am not sure this heading match the related content.
14) About the acquisition date of SAR image in figure 7, there is no match between main text and caption. Check.
15) Lines 505-506: Please clarify point 3.
16) Lines 514-515: since you identify this issue in segmentation, why did you not applied this filtering to mitigate the issue?
17) Please add information of the processing time for all the considered method, not only yours, for comparison.
18) Figure 7: Geographical coordinates should be added. In (a), open water seems to be coded with cyan rather than blue. Is (b) an RGB image or not?
19) Headings related to 5.2 and 5.3 subsections seem to be inverted.
20) Improve references. The format must be unified as well as some references (see [16], [20], [33], [36] and [37]) need to be completed with missing info.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 5 Report
The manuscript addresses an interesting problem, the author can consider giving the manuscript a language/spell check.
Author Response
Dear Reviewer:
Re: Manuscript ID: remotesensing-1973446 and Title: E-MPSPNet: Ice-water SAR Scene Segmentation based on Multi-scale Semantic Features and Edge Supervision
We appreciate the reviewer’s positive evaluation of our work.
We have carefully considered all the comments and revised our manuscript accordingly. In the revised manuscript, the typos and grammar errors we found have been corrected, and the main revisions have been highlighted in red.
We hope our revised manuscript has well addressed your concerns and it can be accepted for publication.
Sincerely.
Round 2
Reviewer 1 Report
Please see the attahced pdf file.
Comments for author File: Comments.pdf
Author Response
Dear Reviewer:
Re: Manuscript ID: remotesensing-1973446 and Title: Ice-water SAR Scene Segmentation based on Multi-scale Semantic Features and Edge Supervision
We have revised our manuscript accordingly. In the revised manuscript, the typos and grammar errors we found have been corrected.
We thank you for the time and effort that you have put into reviewing the previous version of the manuscript.
Sincerely.
Reviewer 2 Report
Thank you for incorporating the changes recommended by the reviewers, and explaining the technique in more detail. The text is much improved, with only minor adjustments needed.
L11-14 In the new abstract text the first 2 sentences read awkwardly and could be rephrased.
L36 This is now missing the icebergs. This should explain that the unpredictable risks to navigation are caused by more icebergs calving as a result of the glacier fronts becoming more exposed.
L79 Replace the "is" in "datasets is complicated" with "being".
L504 Replace "of around the Greenland east" with "from the Greenland Sea".
L505 Replace "which data are" with "containing data that had".
L521 Delete "configured in this paper".
L523 and Figure. Replace "parameters size" with "memory usage".
L622 "Radarsat" should be capitalized, i.e. "RADARSAT".
Author Response
Dear Reviewer:
Re: Manuscript ID: remotesensing-1973446 and Title: Ice-water SAR Scene Segmentation based on Multi-scale Semantic Features and Edge Supervision
We have carefully considered all the comments and revised our manuscript accordingly. In the revised manuscript, the typos and grammar errors we found have been corrected, and the main revisions have been highlighted in red.
Point 8:L523 and Figure. Replace "parameters size" with "memory usage".
Response 8:We are grateful for the suggestion. We have compared the parameter information of the network model . Memory usage refers to the memory size consumed in the training of the model. The size of the model is generally measured by the number of parameters,we consider that “parameters size” is more appropriate than “memory usage”.
Regarding the following minor comments with the question numbers of Point 1-7, we have corrected them. Thank you again for your careful check.
- L11-14 In the new abstract text the first 2 sentences read awkwardly and could be rephrased.
- L36 This is now missing the icebergs. This should explain that the unpredictable risks to navigation are caused by more icebergs calving as a result of the glacier fronts becoming more exposed.
- L79 Replace the "is" in "datasets is complicated" with "being".
- L504 Replace "of around the Greenland east" with "from the Greenland Sea".
- L505 Replace "which data are" with "containing data that had".
- L521 Delete "configured in this paper".
- L622 "Radarsat" should be capitalized, i.e. "RADARSAT".
We thank you for the time and effort that you have put into reviewing the previous version of the manuscript.
Sincerely.
Reviewer 4 Report
All my comments were properly addressed in the new version of the manuscript, which was remarkably improved with respect to the first version. Accordingly, I recommend to publish the paper as it is.
Author Response
Dear Reviewer:
Re: Manuscript ID: remotesensing-1973446 and Title: Ice-water SAR Scene Segmentation based on Multi-scale Semantic Features and Edge Supervision
We appreciate the reviewer’s positive evaluation of our work.We thank you for the time and effort that you have put into reviewing the previous version of the manuscript.
Sincerely.
