Hyperspectral Pansharpening Based on Spectral Constrained Adversarial Autoencoder

Round 1

Reviewer 1 Report

Paper:   remotesensing-622544

Title:      Hyperspectral Pansharpening Based on Spectral Constrained Adversarial Autoencoder

 

This work is concerned with improving the resolution of hyperspectral images in both spectral and spatial domains. The materials were presented in detail. The development, experiment, and results were also given adequately.

However, it is not clear how the network structure, as described in Section 4.2, was determined. In particular, what are the justifications to adopt, for example, 500 hidden layers instead of other numbers? A theoretical support is needed. Otherwise, this work is no difference from any others, and the scientific contribution is not identifiable.

Author Response

Dear Editor and Reviewers,
Thank you very much for patiently handling, giving helpful comments, and thoughtful review of our manuscript entitled "Hyperspectral Pansharpening Based on Spectral Constrained Adversarial Autoencoder" (ID: 622544). The comments and suggestions are all outstanding and valuable, which benefit us a lot in revising and improving our work as well as the essential guiding significance to our further researches. All the authors have seriously discussed all these comments and have tried best to address them and modify the manuscript point by point as follows to meet with the requirements of the journal Remote Sensing.
The revised portions are highlighted in yellow and displayed in the blue font for Abstract in the revised manuscript. We hope that the reviewers could be satisfied with the revisions and responses.
Kindest regards and thanks,
The authors

Comment and suggestions for authors:
This work is concerned with improving the resolution of hyperspectral images in both spectral and spatial domains. The materials were presented in detail. The development, experiment, and results were also given adequately.
Response:
Many thanks for the reviewer's comments for our work. We appreciate the reviewer's spending time and effort in reviewing our paper.
Comment (1):

However, it is not clear how the network structure, as described in Section 4.2, was determined. In particular, what are the justifications to adopt, for example, 500 hidden layers instead of other numbers? A theoretical support is needed.
Response:
We appreciate your valuable comments. We agree with your viewpoint that the network structure is not described sufficiently. We have revised Section 4.2 by adding some interpretations of the parameter settings to explain the network structure further. The number of the hidden nodes in the first two layers is set to 500 to meet the requirement of both the abundance of the spectral information and the performance of the whole network. Additionally, as suggested, we have reported the influence of different numbers of nodes in the last hidden layer for the pansharpening performance in the new Figure (3a) and Figure (3b). It can be observed from the results that the proposed method achieves better pansharpening performance when the number of hidden nodes in the third layer is 30 for almost all the test data sets. Thus, the number of the nodes in the last hidden layer is set to 30. What is more, the influence of depth is also illustrated in the new Figure (3c) and (3d). From the curves of CC and SAM, which has the optimal value of 1 and 0, respectively, we can see that for all the data sets, the pansharpening performance achieves the best when the depth is set to 3.

Comment (2):
Otherwise, this work is no difference from any others, and the scientific contribution is not identifiable.
Response:
We thank the reviewer for the insightful comment. We agree with your opinion that the scientific contribution is not presented precisely enough. Thus, we further supplement the comparison between the proposed method and the existing method. The main innovation of this work is proposing a new hyperspectral pansharpening method based on SCAAE for the first time. The proposed method extracts a deep feature of the HS image as spatial information, improving the performance of the existing shallow-feature-extracting methods, such as PCA based approach. Different from all the existing methods, the proposed method considers both the spatial information of the HS image and PAN image, improving the ability of spatial information enhancement and spectral information preservation. In order to illustrate the innovation and effectiveness of this method objectively, we have added the experiment and description of the leading innovation modules in Section 4.3 Component Analysis. From Table 1 in Section 4.3, we can conclude that compared with the representative dimension reduction approach PCA, the SCAAE based method presents more satisfactory experimental results, which could be identifiable from the other pansharpening methods.

In conclusion, special thanks to you for your valuable comments about the central part of the innovation of our work. The manuscript is revised by expanding the experiment section to verify the contribution of some significant components. We appreciate your warm work earnestly and hope that the correction will meet with approval. Once again, thank you sincerely for your valuable comments and suggestions. We hope everything goes well and looking forward to hearing from you.

Reviewer 2 Report

[Fig. 5] provides the false color pallet [0.0; 1.0]. However, it is not clear, how normalization is obtained. Is it calculated for each sample independently or the global MIN and MAX were applied?

[Fig.7, 9, 11] must have false color pallet too to satisfy one style.

[Fig 11] is provided in conclusions, while it is related with previous chapter.

[Abstract] must provide some quantitative results obtained by authors in the experiment.

[Conclusion] must contain some quantitative results obtained in the experiment and description of possible improvements of experiment or pansharpening method.

[Fig 1] only provides visual effect, but it does not simplify construction of system. The depicted elements are hardly related with sections of their description.

[row 256] training with or without early break by validation dataset? Interruption condition?

[Experimental Setup] the autoencoder is applied. However, it is not clear, is the similar architecture of autoencoder applied for all datasets? (number of layers, activation functions, etc.) Is training augmentation applied? Type?

What is the proportion among training, validation and test datasets?

[Introduction] Is it correct, that architecture of source [61] is applied? If it is true, the SCAEE is not novel method (firstly it is depicted in [61]), the proposed research only improves or validates SCAEE method. Then [rows 83-108] must be precised and rewritten.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

 

GENERAL COMMENTS

The paper presents a new HS pansharpening algorithm based on SCAAE to improve the spatial resolution of LR HS images with HR PAN images. This is an interesting and current research topic in the field – mainly concerning the sensor fusion to automatic assessment and monitoring based on image analysis. Further, the topic is within the journal scope.

In general, the organization of the paper and the sequence of reasoning, in each sub-sections, are clear and enables to understand all the procedure – scope, method and results. The 1. Introduction and ‘2. Related Work’ helps to frame the research presented and, together with the Reference section, it gives a good idea of the state-of-the-art. The propose method, as well as its validation and discussion (based in the experimental results) is well presented and detailed.

ORIENTED COMMENTS

(pg. 3, ln. 91-92) “Compared with the state-of-the-art methods, the proposed method can preserve the spatial and spectral information better.” 'better' is a vague term for a scientific paper... if is not to quantify maybe it’s more appropriate to use 'improve'… (pg.4, ln. 129-) “In this section, firstly, the frequently used methods for HS pansharpening are reviewed, and their existing challenges are analyzed. Then, the proposed method is described in three parts” You should place this section elsewhere, not in the Proposed Method section - I suggest in a subsection of Section 2. Proposed Method - I suggest a final subsection explaining which parameters to use to evaluate and compare the results of the proposed method. Conclusions – The conclusions section needs significant revision. It’s too short and without relevant information.

In the conclusions, it should be clear whether and how the proposed objectives were achieved; and what are the main advantages of the proposed method over the state of the art.

The conclusions should also be supported by the data from Section 4. In this sense it could be useful to summarize the results obtained (can be done at the end of Section 4, and use this summary as a support for the conclusions.

CONCLUSIONS

The manuscript requires major revision.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf