A New Dual-Branch Embedded Multivariate Attention Network for Hyperspectral Remote Sensing Classification

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a new dual-branch network embedded multivariate attention for hyperspectral image classification. Overall, the paper is of scientific sounds. Before possible publication, the following issues must be addressed.

1. The abstract should include the main results and findings, highlighting the core content of the study. For example, the increase of the classification accuracy should be added.

2. The introduction is ambiguous. The author discusses many previous works, however, some of these works are not directly relevant to this paper. It is suggested that the author introduces works most closely related to this paper, followed by outlining the main idea of the method proposed in this paper. Other literature reviews can be separately placed in the section on Related Work.

3. Introduction: Starting from line 130, the authors provide detailed steps of the proposed DMAN. Generally, detailed descriptions should be placed in the Methods section. The reviewer suggests that the authors analyze the motivation behind the proposed method at this point.

4. Figure 1, Why was the threshold for the number of principal components chosen to be 5? Please provide the rationale. Additionally, is this value applicable to all datasets? Typically, for PCA, we choose to retain a certain percentage of the data to determine the threshold.

5. Line 201, characters in " k × k × n" should be italicized.

6. Table 4. we know that the value of kappa coefficient ranges between 0 and 1. However, the values of kappa in Table 4 are all much larger than 1. So, "Kappa" in Table 4 should be written as "Kappa × 100".

7. The authors claim that their method can address small sample data. However, in the experiment, the smallest training set comprises 3% of the entire labeled samples. Therefore, it is suggested that the authors conduct experiments on even smaller training sets. For instance, they could select only a few samples from each class to form the training set.

8. Figure 9 is blurry, its resolution needs to be enhanced.

Author Response

Thank you for your questions and suggestions, our responses to your responses are explained in the document

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This study focuses on A New Dual-Branch Network Embedded Multivariate Attention for Hyperspectral Remote Sensing Classification. The authors provide valuable information on the topic. Here are some general comments for authors to consider.

(1) Please include your numerical results in the abstract. 

(2) Line 180, please rephrase "to act as a classifier, without".

(3) There are several dimension reduction methods, why did the authors use PCA? 

(4) Please evaluate the noise robustness of the algorithm. 

(5) Please compare the effectiveness of the proposed model with other state-of-the-art deep learning models.

(6) Please include the class names in Tables 6,7,8. 

(7) Please run the classification models several times and calculate the standard deviation of the classification results. 

Comments on the Quality of English Language

 Moderate editing of English language required. 

Author Response

Thank you for your questions and suggestions, our responses to your responses are explained in the document

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

1. The abbreviations in Figure 1 should be introduced in the legend or nearby text description.

2. It is suggested that the design motivation and ideas of the method be introduced at the beginning of the Method section to help readers better understand the relationship between the proposed methods.

3. On page four, line 184, the author mentions that few scholars consider using PCA in classification tasks. So, what is the purpose of the author's use of PCA?

4. The paper needs to explain more deeply why the Hybrid Random Patch Model can retain helpful information and why it uses non-learnable convolution models instead of learnable ones. Additionally, can the Hybrid Random Patch Model be transferred to other hyperspectral classification methods to improve performance in a generalized manner?

5. The idea of using the learnable standard deviation of the BN layer to measure channel importance in NAM has been mentioned in other papers. If the author borrowed this idea, it should be cited in the text.

6. In the experiments, the number of training data samples for each class to some extent reflects the proportion of different class samples in the test data. Does this experimental setting allow the model to estimate the proportion of different class samples in the test data through the training data?

7. To demonstrate the superiority of the method, it is recommended to compare it with more state-of-the-art methods. 

Author Response

Thank you for your questions and suggestions, our responses to your responses are explained in the document

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed all my concerns and I suggest to accept the papaer for publication.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors, 

Thank you for the additional review comments.

Best regards, 

Reviewer 

Reviewer 3 Report

Comments and Suggestions for Authors

The response answers my questions.