Attention-Enhanced Generative Adversarial Network for Hyperspectral Imagery Spatial Super-Resolution

Round 1

Reviewer 1 Report

The abstract is generally written without highlighting the main results and conclusions obtained from this study. Hence need to be corrected and reduction is also required at several instances as the sentences are robust.
Introduction is robust.
The problem definition is not defined and identified in the introduction section.
Objectives are defined and, however needs reconstruction for refinement.
The description of the methodology is adequate. Also, the relevant equations for the topic need to be framed theoretically.
The results are well derived and shows good impact on the work. However, there are several laggings in the discussion section.
The discussion of figures 8 and 9 needs to be reconstructed based on the literature and that observed from the figures.
Seems there are several figures and hence recommended to reduce the figure count.
Moreover, the discussion is not consistent for the data presented or figures designed
The conclusions are not consistent with the results. The conclusions should be refined and well structured based on the obtained results.

Moderate English revision is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors propose an attention-enhanced GAN for hyperspectral super-resolution. The proposed method is GAN-based. Moreover, the idea of combining GAN and attention modules is also applied in this method. The main contribution of this paper is that the authors have added two attention modules, spectral attention module (RSAM) and enhanced spatial attention module (ESAM). So the novelty of this paper is limited. The architecture of the proposed method is quite similar to some published works, such as “X. Wang, L. Xie, C. Dong and Y. Shan, "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data," 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), Montreal, BC, Canada, 2021, pp. 1905-1914, doi: 10.1109/ICCVW54120.2021.00217.” The only difference is that the authors use the spectral characteristics of the hyperspectral data. So, I do not recommend to publish this work.

The followings are some detailed comments:

1. The article mainly focuses on the spatial attention mechanism and spectral attention mechanism as innovation points. So they should compare with methods with other attention mechanisms to demonstrate the superiority of the attention mechanisms mentioned in the article?

2. Are the LReLU and the ReLU in line 287 the same activation function? The use of LReLU is not shown in Figure 3.

3. For the total loss function in Eq. (12), it includes too many terms. How to determine the best parameter setting?

4. what is the meaning of the arrows in Table 1-4? And the arrows should be consistent in Table 5 and Table6.

5.In page 4, the authors claimed that "However, the existing GAN based deep learning frameworks for HSI SR frequently suffers from training difficulty, as well as the lack

of further exploration on spatial and spectral contextual information, leading to spatial spectral distortion". Since the proposed method is also GAN-Based, does their method have the same problems. If not, why?

6.In page 7, the authors claimed that "the consumption of GPU memory and other computing resources can be greatly reduced". I suggest they should present a quantitative comparison result.

The English can be improved.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper provides a detailed description of interesting research and a long and decent popular HSI benchmark case study.
Still, there are some small issues that could be improved or supplemented.

The examples provided in Fig.2. are attractive, but seem to be not in the best possible resolution. If possible generate graphics with better resolution of images. If this is an effect of decreased quality while PDF generation, try to use hi-res images in the source document.

Besides the noise and degradation, there is a huge difference in brightness between (a) and (b) images in Fig.2. but it was not discussed. Please provide an extended explanation for that effect.

The MRMSE metric is not defined in the paper context.
There are no comments on why those and no other metrics have been selected.

In the ablation study (4.3), some results are listed, but there is a lack of a straight assessment of the results. To be specific, what is your opinion of the results, do you consider the results significant, promising, or disappointing?

In the third section of the examples presented in Fig 9 (c/f) there are 2 prominent peaks and gaps. As it is present also in the proposed method recovered spec curves it would benefit from your comment. If those are present in the original data it would be also worth it if the reconstruction follows this pattern.

The information on the analyzed data besides the spectral range should be supplemented with the width of provided bands.

One additional minor remark: please consider decreasing the padding in Tab.1. as it exceeds the margins, or transpose the results table.

Finally, such an interesting study should be finalized with further work plan, and propose some following steps that could be considered.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

In the study, the presentation and content of the main contributions part is successful. It is good to implement the experiments with the proposed method on a dataset consisting of 3 open shared widely used hyperspectral images. It is also appropriate to compare the proposed method with the results of state-of-the-art models. My suggestion is that more explanation can be added as to why they were chosen as evaluation metrics (MPSNR etc.) (Because some other more classical metrics are still widely used in studies in this field: F1 score, precision etc.)

The explanation of Table 7 is not clear enough. It is not clear in which dataset it corresponds to the timing. And then, "We can see that our proposed method has the smallest computation burden." There is an explanation. However, when the AEGAN method is compared as inference time in the values in the table, it has an average value. Maybe in this case the comment on 585 can be removed.

In the conclusion part, very few details of the results can be given. It has been said that the proposed method of working in a very general (not exhaustive) way has been successful. Maybe the numerical values of the important part can be shared. Or the amount of improvement it provides can be included.

Instead of the end-to-end "mapping" expression in 211, a different expression may be preferred (only suggestion).

There are typos of English words in some parts of the Manuscript. These errors must be fixed.

Minor editing of English language required.

Round 2

Reviewer 2 Report

The innovation of this paper is low. I do not this kind of work has contributions to the field of remote sensing application.

The English is fine.