Synthesis of Synthetic Hyperspectral Images with Controllable Spectral Variability Using a Generative Adversarial Network
Round 1
Reviewer 1 Report
This paper proposes an GAN based approach to generate synthetic hyperspectral images. By including TV regularization and gradient penalty, the variantion factor can be controlled. Some experiment results are shown to show the effectiveness of the parameters. However, this paper does not give sufficient results to prove the performance of the method.
1. The method proposed in this article produces an effect similar to adding system errors on a spectral curve. Please explain how the method used in this article differs from the data augmentation effect generated by directly multiplying different response curves on the data.
2. Please show the generated sythetic hyperspectral image（in the form of pseudo-color and spectral curves）
3. Please show the comparison results of the proposed method and other methods. Including both traditional methods and neural network methods.
4. Please consider using an application, i.e. classification or detection, to demonstrate the effectiveness of the proposed method.
Author Response
We thank the reviewer for his comment, please find detailed answer in attached file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
See attached file.
Comments for author File: Comments.pdf
Please read my review report properly.
Author Response
We thank the reviewer for his comments, please see detailed answers in the attached file
Author Response File: Author Response.pdf
Reviewer 3 Report
Review Report
The manuscript discusses the application of VAE-GAN in synthesizing hyperspectral images, which is an important topic in image processing and remote sensing. The topics and techniques adopted are relevant in the field, and are widely used in Applied Math and handling remotely sensed images. Nevertheless, the authors have not clearly explained many methodologies, for example, the mathematical or probabilistic transformation needed, denoising process, the use of the generator and/or discriminator etc. (some specific comments and ideas are provided below). The conclusion summarizes the main ideas of the entire manuscript, however it would be better if some potential extension or future research goals can be listed in the manuscript. The references are mostly related to the topic, however, some key references (e.g., applications of VAEs) are missing, the authors should discuss the recent applications of such method before beginning the study.
There are also some significant room for enhancement, which are stated as follows:
(1) Eq. (1): What is the characteristics of the matrix A? What's the relation with the HSI Y?
(2) Lines 30-31: The understanding here might be wrong. One can control the complexity of the model used for generating datasets.
(3) Introduction: The authors should should discuss the applications of VAEs on image processing, image analysis, game design etc. as well, those are possible applications where new samples and new data distribution might be needed. Some references that should be added are as follows:
https://www.mdpi.com/1424-8220/23/7/3457
https://openaccess.thecvf.com/content/CVPR2021/papers/Peng_Generating_Diverse_Structure_for_Image_Inpainting_With_Hierarchical_VQ-VAE_CVPR_2021_paper.pdf
https://www.mdpi.com/2313-433X/7/5/83
(4) Lines 44-45: How? Some specific type of transform will be needed for transforming into a simple distribution..Please clarify.
(5) Lines 39-49: The authors should describe more on denoising process, especially the type of denoising approach adopted in this study.
(6) Line 63: "it" refers to "latent codes"? What are "latent codes" in specific in your study?
(7) Lines 60-63: What actually, or how actually the generator and discriminator help in your current study? What are they in physical manner?
(8) Line 69: "outputting a score instead of a probability" - Why? Need to explain some more details and processing.
(9) Lines 72-74: What's the purpose of smoothen the input signal? Denoising filter might have already done the job once the statistical distribution of the noise components is known.
(10) Eq. (4): Can this 2-norm be replaced by other norms?
(11) Lines 88-90: Why average? What if there are very high or very low values?
(12) Section 3.1 (lines 100-103): Are these parameters optimal? Have any adjustment and trial been conducted?
(13) Lines 105-113: Describe clearly the NMF method, and the pros and cons, as well as the applicability of these 2 methods.
(14) Figure 3: What's the main purpose of setting truncation value and purity?
(15) Lines 127-128, or Section 3: What does truncation value mean or represent in practical daily life?
(16) Lines 129-130: Why stop at truncation of 0.4 and purity of 0.75? i.e., why without continuously increasing the values and test?
Grammatical Typos/ Errors
Line 28: investigate
Line 63: maximizes
Line 2 of Page 3: gets --> receives; stabilize --> stabilizing
Line 75: that we condition,
Line 86: construct
Please conduct a grammatical check before re-submission.
Author Response
We thank the reviewer for his comments, please find detailed response in the attached file
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have replied to all of my comments. I have no further questions.
Author Response
We thank the reviewer for his comments.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors answered all my questions.
Author Response
We thank the review for his comments
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors have tried improving the quality of the manuscript, however some key questions have not yet been addressed, and some references should be added based on the newly added statements.
Comment 2 of previous report: The authors have provided an explanation and more details in their response sheet, however these important ideas should also be incorporated into the revised manuscript.
Comment 3 of previous report: Please add back the references into Lines 44-45 of the new manuscript accordingly. They are quite necessary.
Comment 4 of previous report: Thanks the authors for clarifying it - please provide more details about the forward diffusion process, and why it is used (i.e., the rationale behind).
Comment 6 of previous report: For latent codes of the current study, please include more technical explanations (basically in the response sheet) in the manuscript.
Comment 9 of previous report: "This is a hyperparameter that needs to be tuned once. It seems the same value is good for all datasets." - the authors mentioned this, has this hyperparameter been validated? Please add relevant supporting documents as reference.
Comment 10 of previous report: I think it would be better if the authors can add a "remark" that 1-norm or other norms (other than 2-norm) might not be appropriate.
Comment 12 of previous report: "The parameters are not optimal, but they are good enough to get good results" - how would the error be / how would the result change when the parameters adopted change a bit? An error analysis should be provided in the manuscript based on this important point.
Comment 15 of previous report: The authors have given clear explanation in the response sheet, but some of these ideas should again be included in the manuscript - about the actual meaning and representation of truncation value.
Overall, an additional proper round of revision will be needed.
Some minor grammatical errors persist, please go through the entire manuscript again and check accordingly. Thanks.
Author Response
We thank the reviewer for his comments
Author Response File: Author Response.pdf
