DSSFN: A Dual-Stream Self-Attention Fusion Network for Effective Hyperspectral Image Classification

Round 1

Reviewer 1 Report

The classification of hyperspectral images using deep neural networks is a widely recognized problem that has been addressed by several researchers. However, the proposed structure holds significant value and relevance for a primary reason: the authors have introduced a dual-stream self-attention fusion network that effectively integrates spectral and spatial information to accomplish the task.

 

Unfortunately, the article suffers from some shortcomings. Firstly, the quality of presentation is low. Some figures and tables are not referenced in the text. Secondly, the section that presents the results lacks a discussion regarding the underlying reasons why other methods have achieved better results in certain classes. Thirdly, The quality of the written English is rather poor and need proofread. In spite of these points, this work is interesting and valuable enough to appear in Remote Sensing journal with adequate significant corrections. Therefore, I would recommend acceptance with major corrections.

Comments for author File: Comments.pdf

From a grammatical standpoint, there are a few minor issues such as missing articles or prepositions, inconsistent verb tenses, and a few instances of awkward phrasing. These can be easily addressed through proofreading and minor revisions to ensure grammatical correctness and clarity. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed a hyperspectral image classification with dual stream self-attention. The experimentdemonstrates the effectiveness. There are some questions before acceptance.

1. There is a band selection in the proposed method, how about dimensiona reduction such as PCA? Some abaltion study can be considered to analyze it.

2. The classification map on whole image can be given, instead of only the labeled pixels.

3. Some visualization of the learned features can be given.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In the abstract (lines 9-26), while you clearly outline the overall approach and the general concept of the proposed DSSFN, the technical details regarding the approach are somewhat vague. It would be beneficial if you could provide more precise details about the DSSFN. For instance, how does the self-attention mechanism aid in the extraction of global information? How does the pyramidal residual convolution module reduce computational costs and parameters?

 

The introduction and the abstract seem to repeat some information. It would be more beneficial if you could introduce new insights in the introduction, perhaps including more context or detailing the need for such a study in the current research environment. For instance, the issue regarding the limitations of convolutional kernels being too narrow (line 13-14) is only briefly mentioned in the abstract, but not thoroughly discussed or contextualized in the introduction.

 

While you make a concerted effort to discuss other research and how it relates to your work (lines 59-122), I would suggest that you provide a clear comparison table or visual summary. This could greatly assist readers in understanding the current research landscape, the uniqueness of your approach, and its potential benefits over existing methods.

 

The term "dual-stream" (lines 30-34) needs a proper definition before being used in the context of hyperspectral image classification. Although you have provided references, the direct explanation within the text will make your manuscript more accessible to readers who may be unfamiliar with the concept.

 

Your methodology (lines 104-143) is described broadly and in a fragmented manner. I recommend that you consolidate this section into a more cohesive and clear explanation. Additionally, consider using subheadings to improve the organization and readability of the text. For example, one subheading could be "Feature Extraction," followed by "Band Selection."

 

Be sure to explain all specific terminology and abbreviations when they first appear. For instance, the term "MF-based band sorting strategy" (line 141) should be fully explained or referenced to assist the reader's understanding.

 

The description of the basic approach (lines 154-182) could be improved by clearly delineating between past approaches and how your approach is different or better. This could also be helped with the use of diagrams or flowcharts to visually represent these differences.

 

It would be helpful to include more information about the datasets used in the experiments. Specifically, what types of classes or features are included in these datasets? This information would give readers a better understanding of the complexity of the datasets and how well-suited they are for hyperspectral image classification.

 

While the paper lists the size of the spatial input, depth of the convolutional block, percentage of training set samples, number of spatial self-attention feature extraction layers, and learning rate as important parameters, it does not provide specific details or reasoning behind the chosen values for these parameters. For each parameter, please provide a more detailed explanation of why certain values were chosen and how they affected the model performance.

 

It is noted in section 4.2 that "for each combination of the aforementioned hyperparameters, the model with the best classification performance on the test set is retained on each dataset for comparison with other experimental methods." However, it is not clear whether any form of cross-validation or other robust methods were used to avoid overfitting. Please elaborate on how the model validation was conducted.

 

While the comparisons with contemporary leading approaches are helpful, it would be even more beneficial if the paper could include an analysis of why the proposed DSSFN model outperforms these methods. What specific aspects of the DSSFN model allow it to achieve superior performance?

 

In the experiment results section, you refer to the "optical and spatial patches" used in the DSSFN method. However, these are not described in detail in the provided manuscript section. Please add a detailed explanation or refer to the section of the paper where these patches are explained.

 

It is stated that the DSSFN model is more robust to sparse data, but there is no evidence provided to support this claim. Please provide additional evidence or further explanation to substantiate this claim.

 

For all experimental scenarios, it would be beneficial to include more statistical analysis to evaluate the significance of the results. For instance, hypothesis testing methods can be used to determine whether the DSSFN model's performance improvement is statistically significant.

 

The graphics and figures referred to in the text are not included in this manuscript portion. However, to make your results clearer and more digestible, ensure that your figures clearly illustrate the performance of the DSSFN model compared to the other models and clearly highlight the key findings. Each figure should also have a descriptive caption, and be properly referenced within the text.

 

The paper does well to include the overall accuracy, average accuracy, and kappa coefficient as measures of the model's performance. However, it could be useful to include additional metrics, such as precision, recall, and F1 score. These metrics could provide a more comprehensive view of the model's performance.

 

It is suggested that the proposed method reduces the "misunderstanding between classes 8 and 15". However, it is not clear what the specific characteristics of these classes are that lead to such confusions. An explanation about the specific challenges with these classes would clarify this statement.

Several areas require attention for correction in the paper.

 

Currently, numerous paper sections are composed in colloquial language, which may give the impression of lacking expertise.

 

In my opinion, a more suitable title for Section 2 could be "Preliminaries" or an alternative, instead of "The Basic Approach".

 

It would be beneficial to modify the usage of "Where" in line 302.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Mr. Branislav Dinić

I hope this letter finds you well. I am writing to convey my decision regarding the revised manuscript titled "DSSFN: A Dual-Stream Self-attention Fusion Network for Effective Hyperspectral Image Classification,". I had the privilege of reviewing this article as a referee, and after careful consideration of the revised version, I am pleased to accept the paper for publication in Remote Sensing journal.

Throughout the review process, I had the opportunity to engage in constructive discussions with the authors. They demonstrated commendable responsiveness and addressed the concerns raised during the initial review effectively. I am satisfied with the improvements made to the manuscript, which have significantly enhanced the clarity, quality, and contributions of the research presented.

The authors' work on the proposed "DSSFN" model for hyperspectral image classification, which incorporates a Dual-Stream Self-attention Fusion Network, showcases a novel and promising approach. The experiments conducted and the results obtained underscore the potential impact of this research on the field of hyperspectral imaging.

I believe that the revised version of the article now meets the high standards set by Remote Sensing and is worthy of publication. The authors have diligently addressed the comments and suggestions, making this a valuable addition to the scientific literature.

Once again, I would like to express my gratitude for the opportunity to review this manuscript. It was an enriching experience, and I am confident that this work will make a significant contribution to the scientific community.

Thank you for your time and consideration.

 

 

 

Reviewer 2 Report

no more questions. 

Reviewer 3 Report

I found the authors' overall reflection of my opinion to be quite accurate. While there were a few comments that weren't fully addressed, the authors' explanations did manage to somewhat convince me. I would like to point out a minor issue, however, regarding the capitalization of the word "where" in lines 325 and 511. It should be changed to lowercase "w".

Regarding the capitalization of the word "where" in lines 325 and 511. It should be changed to lowercase "w".