A GAN-Based Augmentation Scheme for SAR Deceptive Jamming Templates with Shadows

Round 1

Reviewer 1 Report

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see the attached document.

Comments for author File: Comments.pdf

Overall, looks good. Minor editing is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Please see the attached file

Comments for author File: Comments.pdf

There are some grammar errors in this paper. Please read carefully to revise them.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

This paper presents a generative network designed to support the development of SAR deception jamming templates with speckle noise and shadows. I have some primary concerns, as outlined below:

 

It is worth considering that existing deep learning models are generally proficient at generating images with minimal speckle noise. Could you please clarify why you introduced a separate noise layer and subsequently incorporated it into structural data?

 

The training process appears to lack crucial information, such as the loss variation curve over the course of training.

 

In lines 250-251, you mentioned, "The MSTAR data were collected using the Sandia National Laboratories SAR sensor platform and the Sentinel-1A multimode SAR sensor platform with X-band SAR sensors." Could you verify the accuracy of this statement?

 

Regarding lines 318-320, could you provide insights into how you determined the similarity (75.44%) based on ENL values? The same query applies to lines 324-325. I believe that relying solely on ENL and SSIM, which are based on image brightness statistics and do not account for geometric structures, might not be sufficient to validate performance.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revised paper corrects the grammatical errors in the original paper, improves the clarity of the figures, and supplements some of the experimental details, but it still has the following serious deficiencies: (1) The evaluation criterion of the deceptive jamming template is authenticity, which includes both the authenticity of speckle noise and the authentic correspondence between the shadow region and the target. But the evaluation indexes ENL, AG, MSD, SSIM used in the paper can only evaluate the similarity between the generated image and the original image, and cannot evaluate whether the generated speckle noise is authentic or not, and whether the shadow region meets the authentic correspondence with the target. If the similarity tends to 1 as the evaluation criterion of high-quality SAR image enhancement model, it is not necessary to train the GAN model, just copy the source image. (2) In paper L310-311, "When the number of training times is 2000, the value of the loss function is close to 0, indicating that the model performs well". The loss function of the single-image-based generative model is close to 0, indicating precisely that the training overfits. When the loss function is equal to 0, the generated image is the source image. (3) Lack of comparison with existing deep learning-based sample augmentation schemes for deception jamming templates

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Thank you for considering my comments,

Below are my comments on the revised version:

- Figure 4 needs more details about the generator; it gives the impression that the generator is receiving real images as inputs. Is that right?

- The sentence in line 304 is really confusing.

- Figure 6 needs clarification.

- I am concerned about the results presented in Figure 8. Is there any difference between the real and the generated samples? If not, why and how is it useful?

- Captions and figures shouldn't be on different pages (Figure 14).

- Lines 488-493 are confusing or need proofreading.

- All the new content in red needs proofreading.

 

- The discussion section doesn't include the new comparisons.

 

Minor editing is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The author has made a good revision of the paper, and there is a question that needs to be asked: whether the method proposed in this paper can be applied to the SAR ship detection model, and some work can be referred, such as https://doi.org/10.3390/rs13214209.

This paper can be accepted after carefully revision.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

I still don't think you can get a quantitative similarity calculation by comparing the statistical measures of each graph. There are many similarity evaluation criteria between graphs that can be referred to.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf