Hyperspectral Image Classification Based on Fusing S³-PCA, 2D-SSA and Random Patch Network

Round 1

Reviewer 1 Report

The paper proposes a novel HSI classification network based on Random Patches Network (RPNet). The line of research is highly relevant and the proposal is very promising, nevertheless, I recommend the paper to be majorly revised considering the following comments that are mainly related to the technical contribution and the strength of the simulations.

(1)      The technical contribution of the paper is weak and lacks of the theoretical guarantee. Thus, it is necessary to verify that the proposed classification method can reasonably express the mutual encouragement between the S3-PCA and RPNet.

(2)      The section “abstract” is not readily readable, I suggest to rewrite the contributions. As far as know, the network can provide advantage of the data-driven, and PCA scheme can consider the spectral knowledge. Thus, the authors should emphasis the contribution from the physical interpretation rather than only use the CNN framework.

(3)      Besides the classification task, recently some new attention models are proposed for HSI restoration and denoising tasks [1-2]. Thus, the authors should discuss these works about CNN, which can provide some theory supports in terms of contributions. Further, the tensor-based models also are applied to the feature extraction and classification of an HSI as [3-5], thus I suggest the authors to discuss these works.

[1] Variational Regularization Network with Attentive Deep Prior for Hyperspectral–Multispectral Image Fusion, IEEE Transactions on Geoscience and Remote Sensing, 10.1109/TGRS.2021.3080697.

[2] X. Li, M. Ding, Y. Gu and A. Pižurica, An End-to-End Framework for Joint Denoising and Classification of Hyperspectral Images. IEEE Transactions on Neural Networks and Learning Systems.

[4] Multilayer Sparsity-Based Tensor Decomposition for Low-Rank Tensor Completion. IEEE Transactions on Neural Networks and Learning Systems (2021).

[5] H. Zeng, J. Xue, H. Q. Luong and W. Philips, "Multimodal Core Tensor Factorization and Its Applications to Low-Rank Tensor Completion," in IEEE Transactions on Multimedia, 2022.

[5] Unsupervised Deep Tensor Network for Hyperspectral–Multispectral Image Fusion, IEEE TNNLS 2023.

(4)    The authors neglect the analysis of parameter number with the comparison methods.

Moderate editing of English language required

Author Response

Dear reviewer 1#:

   Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. We already carefully read the comments from you and revised manuscript according to your suggestions. We hope that this revision will make our manuscript to meet publish. The responses to the comments point by point are listed as below. Please feel free to contact us with any questions. If the revised manuscript maybe exist the shortcomings, please tell us. We will try our best to continue to revise our manuscript in order to improve our manuscript. Really thank your insightful comments and help again!

Yours sincerely，

  Regards,

  Tao Chen

Reviewer #1:

The paper proposes a novel HSI classification network based on Random Patches Network (RPNet). The line of research is highly relevant and the proposal is very promising, nevertheless, I recommend the paper to be majorly revised considering the following comments that are mainly related to the technical contribution and the strength of the simulations.

• The technical contribution of the paper is weak and lacks of the theoretical guarantee. Thus, it is necessary to verify that the proposed classification method can reasonably express the mutual encouragement between the S³-PCA and RPNet.
• The section “abstract” is not readily readable, I suggest to rewrite the contributions. As far as know, the network can provide advantage of the data-driven, and PCA scheme can consider the spectral knowledge. Thus, the authors should emphasis the contribution from the physical interpretation rather than only use the CNN framework.
• Besides the classification task, recently some new attention models are proposed for HSI restoration and denoising tasks [1-2]. Thus, the authors should discuss these works about CNN, which can provide some theory supports in terms of contributions. Further, the tensor-based models also are applied to the feature extraction and classification of an HSI as [3-5], thus I suggest the authors to discuss these works.

[1] Variational Regularization Network with Attentive Deep Prior for Hyperspectral–Multispectral Image Fusion, IEEE Transactions on Geoscience and Remote Sensing, 10.1109/TGRS.2021.3080697.

[2] X. Li, M. Ding, Y. Gu and A. Pižurica, An End-to-End Framework for Joint Denoising and Classification of Hyperspectral Images. IEEE Transactions on Neural Networks and Learning Systems.

[4] Multilayer Sparsity-Based Tensor Decomposition for Low-Rank Tensor Completion. IEEE Transactions on Neural Networks and Learning Systems (2021).

[5] H. Zeng, J. Xue, H. Q. Luong and W. Philips, "Multimodal Core Tensor Factorization and Its Applications to Low-Rank Tensor Completion," in IEEE Transactions on Multimedia, 2022.

[5] Unsupervised Deep Tensor Network for Hyperspectral–Multispectral Image Fusion, IEEE TNNLS 2023.

(4) The authors neglect the analysis of parameter number with the comparison methods.

COMMENT 1: The technical contribution of the paper is weak and lacks of the theoretical guarantee. Thus, it is necessary to verify that the proposed classification method can reasonably express the mutual encouragement between the S³-PCA and RPNet.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have verified the validity of the S³-PCA and RPNet in Table 5-7. and made some explanation in the discussion section Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 2: The section “abstract” is not readily readable, I suggest to rewrite the contributions. As far as know, the network can provide advantage of the data-driven, and PCA scheme can consider the spectral knowledge. Thus, the authors should emphasis the contribution from the physical interpretation rather than only use the CNN framework.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially rewritten the section “abstract”and emphasized the contribution from the physical interpretation to make readers better understand our article. It has to be said that the expert advice is very correct. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 3: Besides the classification task, recently some new attention models are proposed for HSI restoration and denoising tasks [1-2]. Thus, the authors should discuss these works about CNN,which can provide some theory supports in terms of contributions. Further, the tensor-based models also are applied to the feature extraction and classification of an HSI as [3-5], thus I suggest the authors to discuss these works.

[1] Variational Regularization Network with Attentive Deep Prior for Hyperspectral–Multispectral Image Fusion, IEEE Transactions on Geoscience and Remote Sensing, 10.1109/TGRS.2021.3080697.

[2] X. Li, M. Ding, Y. Gu and A. Pižurica, An End-to-End Framework for Joint Denoising and Classification of Hyperspectral Images. IEEE Transactions on Neural Networks and Learning Systems.

[4] Multilayer Sparsity-Based Tensor Decomposition for Low-Rank Tensor Completion. IEEE Transactions on Neural Networks and Learning Systems (2021).

[5] H. Zeng, J. Xue, H. Q. Luong and W. Philips, "Multimodal Core Tensor Factorization and Its Applications to Low-Rank Tensor Completion," in IEEE Transactions on Multimedia, 2022.

[5] Unsupervised Deep Tensor Network for Hyperspectral–Multispectral Image Fusion, IEEE TNNLS 2023.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have cited more related works in the in the introduction part(lines 92-106), in order to justify our article. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 4: The authors neglect the analysis of parameter number with the comparison methods.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially revised our manuscript in order to clearly describe the analysis in Section 3.2. At present, the Section 3.2 discussed compared the effect of different key parameters. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

And so on, please read our revised manuscript. We thank the comments and opportunity for us to improve our manuscript. As much as possible, the questions were taken into account during the preparation of the revised manuscript. We hope that the manuscript is now suitable for publication.

Author Response File: Author Response.docx

Reviewer 2 Report

1.     The presentation of this manuscript should be revised. Moreover, the writing needs to be improved and carefully revised.

2.     For all equations not authored by the research, it is preferable to refer to their main reference.

3.     In the proposed method, please highlight the novelty.

4.     In the related works, the authors have not covered many recent works in the field of hyperspectral images classification. Recent 2 years, many high-quality methods are proposed, the authors should read them.

5.     Please comment on the complexity of the proposed method.

6.     Experiments are not enough to validate the proposed approach. I strongly suggest the authors compare their method to the current state-of-the-art methods, i.e., works proposed in 2022 and 2023.

7.     In the conclusion section, please add a few sentences talking about the limitations of this study.

Author Response

Dear reviewer 2#:

   Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. We already carefully read the comments from you and revised manuscript according to your suggestions. We hope that this revision will make our manuscript to meet publish. The responses to the comments point by point are listed as below. Please feel free to contact us with any questions. If the revised manuscript maybe exist the shortcomings, please tell us. We will try our best to continue to revise our manuscript in order to improve our manuscript. Really thank your insightful comments and help again!

Yours sincerely，

  Regards,

  Tao Chen

Reviewer #2:

1. The presentation of this manuscript should be revised. Moreover, the writing needs to be improved and carefully revised.
2. For all equations not authored by the research, it is preferable to refer to their main reference.
3. In the proposed method, please highlight the novelty.
4. In the related works, the authors have not covered many recent works in the field of hyperspectral images classification. Recent 2 years, many high-quality methods are proposed, the authors should read them.
5. Please comment on the complexity of the proposed method.
6. Experiments are not enough to validate the proposed approach. I strongly suggest the authors compare their method to the current state-of-the-art methods, i.e., works proposed in 2022 and 2023.
7. In the conclusion section, please add a few sentences talking about the limitations of this study.

COMMENT 1: The presentation of this manuscript should be revised. Moreover, the writing needs to be improved and carefully revised.

RESPONSE: Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. According to expert advice, we have substantially revised our manuscript to make it more logical so that readers can better obtain relevant information. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 2: For all equations not authored by the research, it is preferable to refer to their main reference.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have checked the quoted equations and referred to their main reference. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 3: In the proposed method, please highlight the novelty.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially rewritten the section “abstract”and emphasized the novelty to make readers better understand our article. It has to be said that the expert advice is very correct. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 4:In the related works, the authors have not covered many recent works in the field of hyperspectral images classification. Recent 2 years, many high-quality methods are proposed, the authors should read them.

RESPONSE: Thank you very much for the insightful comments. According to expert advices, we have substantially revised our manuscript in the introduction part(lines 92-106, 115-121) to introduce more relevant theories, techniques and novel structural framework. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 5: Please comment on the complexity of the proposed method.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added the instruction about the complexity of the proposed method and put forward several theoretical statements and propositions in the discussion section (lines 402-409). Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 6: Experiments are not enough to validate the proposed approach. I strongly suggest the authors compare their method to the current state-of-the-art methods, i.e., works proposed in 2022 and 2023.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added three novel methods into the experiment in Tables 5-7 and Figures 12-14. to justify the validity of our method. Moreover, we made some statements in the discussion section (lines 397-402) to explain more.Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 7: In the conclusion section, please add a few sentences talking about the limitations of this study.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added some description about the limitation of our method in the conclusion section (lines 440-444). Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

And so on, please read our revised manuscript. We thank the comments and opportunity for us to improve our manuscript. As much as possible, the questions were taken into account during the preparation of the revised manuscript. We hope that the manuscript is now suitable for publication.

Author Response File: Author Response.docx

Reviewer 3 Report

I consider it important to strengthen the abstract. A clear writing order should be presented for better reader comprehension: general background, experimental approach, high-level description of methods and results, most significant findings, and possible future approaches. The wording of lines 36-36 can be substantially improved. Additionally, a good synthesis of lines 174-184 can complement and greatly strengthen the abstract. Adding a results diagram instead of a table in the discussion section would enhance quick reader understanding. Good job.

Author Response

Dear reviewer 3#:

   Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. We already carefully read the comments from you and revised manuscript according to your suggestions. We hope that this revision will make our manuscript to meet publish. The responses to the comments point by point are listed as below. Please feel free to contact us with any questions. If the revised manuscript maybe exist the shortcomings, please tell us. We will try our best to continue to revise our manuscript in order to improve our manuscript. Really thank your insightful comments and help again!

Yours sincerely，

  Regards,

   Tao Chen

Reviewer #3:

I consider it important to strengthen the abstract. A clear writing order should be presented for better reader comprehension: general background, experimental approach, high-level description of methods and results, most significant findings, and possible future approaches. The wording of lines 36-36 can be substantially improved. Additionally, a good synthesis of lines 174-184 can complement and greatly strengthen the abstract. Adding a results diagram instead of a table in the discussion section would enhance quick reader understanding. Good job.

COMMENT 1: I consider it important to strengthen the abstract. A clear writing order should be presented for better reader comprehension: general background, experimental approach, high-level description of methods and results, most significant findings, and possible future approaches.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially rewritten the section “abstract” to make my article more logical to make readers better understand it. It has to be said that the expert advice is very correct. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

Abstract: Recently, the rapid development of deep learning has greatly improved the performance of image classification. However, a central problem in hyperspectral image (HSI) classification is spectral uncertainty, where spectral features alone cannot accurately and robustly identify a pixel point in a hyperspectral image. This paper presents a novel HSI classification network called MS-RPNet i.e. multiscale superpixelwise RPNet, which combining superpixel-based S3-PCA with two-dimensional singular spectrum analysis(2D-SSA) based on the Random Patches Network (RPNet). The proposed frame can not only take advantage of the data-driven, but also apply S3-PCA to efficiently consider more global and local spectral knowledge at super-pixel level. Meanwhile, 2D-SSA is used to noise removal and spatial feature extraction. Then, the final features are obtained by random patch convolution and other steps according to the cascade structure of RPNet. The layered extraction superimposes the spatial different information into multi-scale spatial features, which complements the feature from various land covers. Finally, the final fusion features were classified by SVM to get the final classification results. The experimental results on several HSI datasets demonstrate the effectiveness and efficiency of MS-RPNet which outperform several current state-of-the-art methods.

COMMENT 2:The wording of lines 36-36 can be substantially improved.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have reorganized my words in lines 39-40 (at present).Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 3:Additionally, a good synthesis of lines 174-184 can complement and greatly strengthen the abstract.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially revised our abstract section referring to make a synthesis of lines you mentioned. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 4:Adding a results diagram instead of a table in the discussion section would enhance quick reader understanding.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added intuitive diagram in Figure 12. after Tables 5-7 for enabling readers to understand our article more clear.Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

And so on, please read our revised manuscript. We thank the comments and opportunity for us to improve our manuscript. As much as possible, the questions were taken into account during the preparation of the revised manuscript. We hope that the manuscript is now suitable for publication.

Author Response File: Author Response.docx

Reviewer 4 Report

In this manuscript, based on fusing spectral-spatial and superPCA (S^3-PCA), two-dimensional singular spectrum analysis (2D-SSA) and random patch network, hyperspectral image classification has been studied. Though most details are presented in the paper, it still needs some revision before publishing.

A few suggestions to improve this paper:

(1) Abstract: The abstract needs to be reorganized and revised, some import results and conclusions in this study needs to be summarized in the abstract.

(2) Line 43-45: The expression of “principal component analysis”, “(Linear Discriminant Analysis” and “maximum noise fraction” do not need to appear twice.

(3) Line 69, 76: The repression of “morphological attribute profiles” and “singular value decomposition” also do not need to appear twice.

(4) Line 135: The font of the formula symbol must be consistent.

(5) Figure 1: The texts in the block diagram “2D-SSA” are not clear.

(6) Line 215: What the meaning of “*” in this equation?

(7) Figure 4: Different from Fig. 2-3, why the first principal component can’t the capture the feature of legend region in ground-truth map, which should be given corresponding explain.

(8) Line 257: “Figure.5” -> “Figure 5”

(9) Fig. 5: This figure is not clear and needs to be improved. Besides, for subfigure (b) and (c), what is the value of PC_num used?

(10) Line 270: Some corresponding explanation need to be added for Fig 6-8.

(11) Conclusions: the part needs further summarized and improved.

(12) The English language also needs to be further improved in this manuscript.

Moderate editing of English language required.

Author Response

Dear reviewer 4#:

   Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. We already carefully read the comments from you and revised manuscript according to your suggestions. We hope that this revision will make our manuscript to meet publish. The responses to the comments point by point are listed as below. Please feel free to contact us with any questions. If the revised manuscript maybe exist the shortcomings, please tell us. We will try our best to continue to revise our manuscript in order to improve our manuscript. Really thank your insightful comments and help again!

Yours sincerely，

  Regards,

  Tao Chen

Reviewer #4:

In this manuscript, based on fusing spectral-spatial and superPCA (S³-PCA), two-dimensional singular spectrum analysis (2D-SSA) and random patch network, hyperspectral image classification has been studied. Though most details are presented in the paper, it still needs some revision before publishing.

A few suggestions to improve this paper:

(1) Abstract: The abstract needs to be reorganized and revised, some import results and conclusions in this study needs to be summarized in the abstract.

(2) Line 43-45: The expression of “principal component analysis”, “(Linear Discriminant Analysis” and “maximum noise fraction” do not need to appear twice.

(3) Line 69, 76: The repression of “morphological attribute profiles” and “singular value decomposition” also do not need to appear twice.

(4) Line 135: The font of the formula symbol must be consistent.

(5) Figure 1: The texts in the block diagram “2D-SSA” are not clear.

(6) Line 215: What the meaning of “*” in this equation?

(7) Figure 4: Different from Fig. 2-3, why the first principal component can’t the capture the feature of legend region in ground-truth map, which should be given corresponding explain.

(8) Line 257: “Figure.5” -> “Figure 5”

(9) Fig. 5: This figure is not clear and needs to be improved. Besides, for subfigure (b) and (c), what is the value of PC_num used?

(10) Line 270: Some corresponding explanation need to be added for Fig 6-8.

(11) Conclusions: the part needs further summarized and improved.

(12) The English language also needs to be further improved in this manuscript.

COMMENT 1: Abstract: The abstract needs to be reorganized and revised, some import results and conclusions in this study needs to be summarized in the abstract.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially rewritten the section “abstract” to make my article more logical and abundant to make readers better understand it. It has to be said that the expert advice is very correct. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 2: - Line 43-45: The expression of “principal component analysis”, “(Linear Discriminant Analysis” and “maximum noise fraction” do not need to appear twice.

- Line 69, 76: The repression of “morphological attribute profiles” and “singular value decomposition” also do not need to appear twice.

- Line 135: The font of the formula symbol must be consistent.

- Figure 1: The texts in the block diagram “2D-SSA” are not clear.

- Line 257: “Figure.5” -> “Figure 5”

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially revised our manuscript in order to to avoid to duplicate sentences and lines in the same paper. In our revised manuscript, we have provided more relevant content to bring theoretical contributions and practical implications into focus. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 3: Line 215: What the meaning of “*” in this equation?

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added some explanation to the symbol(line 254).  Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 4: Figure 4: Different from Fig. 2-3, why the first principal component can’t the capture the feature of legend region in ground-truth map, which should be given corresponding explain.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added the reason and explanation about why the first principal component can’t the capture the feature of legend region in ground-truth map (lines 279-282) to make readers better understand our article. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 5: Fig. 5: This figure is not clear and needs to be improved. Besides, for subfigure (b) and (c), what is the value of PC_num used?

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have replaced the clear picture to make it easier for people to read the information for Figure 5. Then, we have added the information about the set of PC_num (lines 331,359). Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 6: Line 270: Some corresponding explanation need to be added for Fig 6-8.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have added some description and explanation about Figures 6-8 (lines 333-339). Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 7: Conclusions: the part needs further summarized and improved.

RESPONSE: Thank you very much for the insightful comments. According to expert advice, we have substantially revised the Section 5 in order to enrich the conclusion and pointed out the future research direction. In addition, we have added to acknowledge limitations and develop future research lines in the conclusion section (lines 440-444). Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

COMMENT 8: The English language also needs to be further improved in this manuscript.

RESPONSE: Thank you very much for the insightful comments. Thank you for giving us a choice to correcting the shortcoming of our manuscript. According to expert advice, we have substantially revised our manuscript to make it more logical so that readers can better obtain relevant information. Thank you very much for your insightful comment once again. Please read revised manuscript, thanks!

And so on, please read our revised manuscript. We thank the comments and opportunity for us to improve our manuscript. As much as possible, the questions were taken into account during the preparation of the revised manuscript. We hope that the manuscript is now suitable for publication.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have addressed my comments, thus I suggest to accept this manscript for publishment.

Reviewer 2 Report

The authors improved the manuscript according to the comments. Therefore, I recommend acceptance of this version.

Reviewer 4 Report

Since the authors have revised the manuscript carefully by following my suggestion, I don't have other comments.