The Use of a Stable Super-Resolution Generative Adversarial Network (SSRGAN) on Remote Sensing Images

Round 1

Reviewer 1 Report

In this manuscript, the authors proposed a GAN model based on more complicated Generator and Discriminator to achieve an anticipated better performance than the others in Super Resolution. I have some suggestions:

1. In my opinion, to demonstrate the effectiveness of the proposed generator and discriminator. The ablation study has to be performed, which is a critical experiment. 

2. If possible, please show some bad results and try to explain the limitations of the proposed method and analysis. This would improve the quality of this manuscript.

3. I am confused of "Time" in Table 1,2, and 3. Please be specific. Time for training or time for executing.

4. Please try to explain why to use GAN for Super Resolution. What are the advantages of using GAN than the others?

In my opinion, minor editing of English language is required

Author Response

We would like to take this opportunity to thank you for all your time involved and this great opportunity for us to improve the manuscript. We have uploaded our point-by-point response to the comments (response to reviewers). Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Article ID:      remotesensing-2663858-peer-review-v1

Article Title:   The Use of a Stable Super-resolution Generative Adversarial Network (SSRGAN) on Remote Sensing Images

 

 

Summary:

 

The challenging imaging circumstances in orbit and satellites' finite imaging systems limit remote sensing image resolution. Image super-resolution improves image clarity. Motion, imaging blur, down-sampling matrix, and noise impact high-resolution (HR) images. Changes in these conditions severely impact low-resolution (LR) images, making super-resolution reconstruction a pathological anti-problem. We used deep learning to rebuild satellite images to improve imaging quality without modifying the optical equipment. We updated the original super-resolution generative adversarial nets network, upgraded the generator's network to Res-Net-50, and added a fully connected (FC) layer to the discriminator's network. To maintain more detail, we adjusted the loss function by increasing the weight of regularisation loss from 2 × 10−8 to 2 × 10−9. We also carefully selected low-light remote sensing photos from GF-5 satellites to create a fresh dataset for training and validation. The test results showed our strategy works. At scaling factors 2, 3, and 4, the reconstruction peak signal-to-noise (PSNR) was 32.6847, 31.8191, and 30.5095 dB, and the structural similarity (SSIM) was 0.8962, 0.8434, and 0.8124. The super-resolution speed was good, making real-time reconstruction possible.

Therefore, it is interesting and attractive. However, it should be major revised to enhance the quality, as follows:

 

1) In Section 1, authors should make three sub sections, motivation, contributions and organization of the paper.

2) Literature review is not upto the mark. Pl include one table for comparison analysis.

3) A summary table should be provided for convenience for the readers in the literature review section with comparison analysis of other approaches.

4) Contributions of the research paper is not clear, Pl mention in contribution subsection with 3 highlighted points.

5) Overall work methodology.is not clear. Pl elaborate it clearly.

6) Eq 4 and 5 are not clear. Pl represents with specific attributes.

7) Fig 3 ,10 and 11 are not derive properly. Pl derive properly.

8)  Pl mention clearly about the fig 12 ,13 and 14 as part(a) and part(b).

9) Pl compare your work with existing work. It is missing in the present manuscript.

10) Finally, the authors should double-check all formation, typos, and writing throughout the paper.

Author Response

We would like to take this opportunity to thank you for all your time involved and this great opportunity for us to improve the manuscript. We have uploaded our point-by-point response to the comments (response to reviewers). Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

It should be clarified if in order to use the proposed worflow it is necessary to have the high-resolution images, as shown in the proposed tests, or it’s possible to perform upscaling of degraded images whitout the high resolution versions.

1) It would be appropriate to state the pixel resolution of the images used during the training and validation phases of the algorithm;

2) On pg .12, at the conclusion of section 4.2.1. it would be appropriate to state, according to the authors, what are the optimal Time/s parameters for which a reconstruction process can be called efficient;

3) The results of the validation tests are good; however, there is a lack of reflection on future research outlets. Is the process also applicable to already degraded and low-resolution images, or, as in the case of the test, should one always downsample high-definition images?

average

Author Response

We would like to take this opportunity to thank you for all your time involved and this great opportunity for us to improve the manuscript. We have uploaded our point-by-point response to the comments (response to reviewers). Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

After carefully reviewing, the authors have addressed all my questions, and I have no other extended questions.

Reviewer 2 Report

Authors are addressed all the queries related to the present manuscript. Now , it may considered for the publications in this journal.