Super-Resolution for Hyperspectral Remote Sensing Images Based on the 3D Attention-SRGAN Network
Round 1
Reviewer 1 Report
In my opinion, this paper was suitable for publishing before the last changes. After accepting the reviewers' suggestions the paper has been improved.
Author Response
Thank you very much for your comments. We have accpected all of the reviewers' suggestions and revised our manuscript. We hope our manuscript would be suitable for publishing this time.
Reviewer 2 Report
Dear Authors;
I have some comments on your manuscript, I ask you to read it and make the required modifications and corrections.
1- Page 3 of 28, Line 125, you said Section provides a brief introduction to SRGAN, but I have found only a related work on GAN. what about GAN applications? I suggest you to remove Section 2.
2- Page 8 of 28, Line 301, could you give some references for this sentence (should be trained independently at first to avoid local minimum) or you have to explain how this can happen?
3- Page 9 of 28, Line 333, c2=(k2L)2.
4- Page 2 of 28, Line 604, it is Table 5 not Table 4, check it please.
5- I think you can remove the paragraph from Conclusion section started at line 623 (future studies can focus on ...) to the end of this section.
6- I suggest you write some results at Conclusion section.
Author Response
Response to Reviewer’s Comments
Point 1: Page 3 of 28, Line 125, you said Section provides a brief introduction to SRGAN, but I have found only a related work on GAN. what about GAN applications? I suggest you to remove Section 2.
Response 1: We changed the structure of our article by deleting Section 2 to make it more consisted.
Point 2: Page 8 of 28, Line 301, could you give some references for this sentence (should be trained independently at first to avoid local minimum) or you have to explain how this can happen?
Response 2: We added the reference for the sentence (Page 8 of 28, line 301, should be trained independently at first to avoid local minimum) to better explain the principle behind it and make it more logical to the readers.
Point 3: Page 9 of 28, Line 333, c2=(k2L)2.
Response 3: We revised the typo in Page 9 of 28, Line 333.
Point 4: Page 2 of 28, Line 604, it is Table 5 not Table 4, check it please.
Response 4: We re-checked the content in Section 4.4 and changed the typo in Page 2 of 28, Line 604 from Table 5 to Table 4.
Point 5: I think you can remove the paragraph from Conclusion section started at line 623 (future studies can focus on ...) to the end of this section.
Response 5: We removed the paragraph from Conclusion section started at line 623 (future studies can focus on ...) to the end of this section according to the comment.
Point 6: I suggest you write some results at Conclusion section.
Response 6: We added some content about the results of our experiment in the Conclusion section to summarize our results.
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
The authors answered correctly to my comments and improved the paper accordingly. I thank them for that. I recommend the revised paper for publication.
Reviewer 2 Report
This paper Super-Resolution for Hyperspectral Remote Sensing Image Using Generative Adversarial Network based on 3D Convolution Network presents an interesting methodology, which has the potential for wider use in the future. The paper is well structured. Introduction has nice literature review; the methodology is well presented.
The application of neutral networks seems OK, maybe more details about relations and characteristic of test/training data would be nice to present. Consider to make a map/scheme which will shot the relation between test and training data, (ln. 329).
Results are well presented and explained. Figures does not show - highlight your results I relations to other results. Figure caption could be a little more extensive in explaining the figure elements.
Overview map/scheme that shows the boundaries of results areas (fig 4-7) would help better present results, and to understand the relations between shown areas. Can you add some scale to presented images (not necessary)?
Results shown on figure 6 are a bit too poorly presented in text.
Discussion and conclusion sections are very well.
Reviewer 3 Report
This reviewer is new for this manuscript and did not consider the response to the review file.
This manuscript proposes to use a 3D GAN network for hyperspectral image super-resolution problem. The experiment section is convincing with expected good results.
The introduction section mentions three novelties.
First: change from 2D to 3D network.
Second: Additional loss term with spectral angle.
Third: Removing the Batch Normalization layer.
Unfortunately, for this reviewer, some more novelties are required for a journal publication. Changing the network architecture to 3D or removing the batch norm may not be considered as novel ideas.
Another primary issue is the writing of the manuscript.
The new paragraph from line 56-73 lists the existing methods without giving any insights/relevance/limitations as compared to the proposed approach.
Further, there is no need for this paragraph from lines 56-73. Instead, the literature review in section 2.1 is more relevant.
Since there is a large number of publications on super-resolution, it is not required to cite so many articles, but maybe differences can be mentioned with the appropriate methods.
Section 2.2 and 2.3 not required.
Section 3.1 describes what 2d and 3d convolution is. Thus can be avoided.
