SAR Image Classification Using Gated Channel Attention Based Convolutional Neural Network

Round 1

Reviewer 1 Report (Previous Reviewer 2)

The authors have addressed my concerns relatively well.

Author Response

Dear Reviewer #1:

 

Thank you again for your valuable suggestions.

We revise the manuscript carefully according to your comments.

The item-by-item response to the critical issues is as follows.

In the new manuscript, revisions are marked with red fonts.

 

• The authors have addressed my concerns relatively well.

Response: Thank you very much for your approval of our responses and revisions. Your comments are so valuable and the quality of our manuscript has been improved obviously according to your comments

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

The paper is well written, with a clear focus on contributions. The results are explained comprehensively. 

Authors are requested to provide comments on the computational complexity of the proposed CNN and compare it with SVM, SAE and DBN.

 

Author Response

Dear Reviewer #2:

 

Thank you again for your valuable suggestions.

We revise the manuscript carefully according to your comments.

The item-by-item response to the critical issues is as follows.

In the new manuscript, revisions are marked with red fonts.

 

• The paper is well written, with a clear focus on contributions. The results are explained comprehensively. .

Response: Thank you very much for your approval of the quality of our manuscript. The manuscript has been revised carefully based on your valuable comments.

 

• Authors are requested to provide comments on the computational complexity of the proposed CNN and compare it with SVM, SAE and DBN.

Response: Thank you very much for this valuable comment. Based on this comment, the computational complexity of the GCA-CNN, CNN, DBN, SAE and the SVM are discussed from the aspects of time complexity and spatial complexity. Section 4.7 are added in the revised manuscript to discuss the complexity of these algorithms. Please see the revised manuscript for the details of the discussion made in section 4.7. In order to make the discussion more readable, the  structural parameters of the DBN and SAE are given in the section 4.2.  

 

 

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

This is a good research article, overall.  

Things I would recommend to improve it include:

For all abbreviations or acronyms, be sure to spell out the first use where it is first used, even if this occurs in the abstract.  Example:  SAR.

Another example occurs on line 117, ReLU.  Later you spell it out.  Should spell it out at first use.  Typical throughout document.

There are some spacing issues that need repair in the abstract.

Keep the titles/headers for Figures and Table with the graphic.  Do not allow it to go to the next page separate from the graphic.

Line 238/239:  The sentence is not clear.  What is meant to be said there?

In Figures 5, 7, and 9:  Spend some time in the document giving thoughts as to why the peaks occur in these figures.

Although your method had the best overall accuracy as compared to other methods, the other methods did beat your method a considerable amount.  For example, in Table 1, for class 1 your method was beat 1 time.  For classes 2 through 8, your method was beat, 3, 2, 3, 0, 3, 0, 2, 0 times.  Thus, there should be some discussion in the document as to why other methods may be better in certain circumstances or, on the contrary, tell why this is not an issue.  In table 2, your method was beat: 3, 1, 1, 1, 0 times for the 5 classes.  In table 3, your method was beat:  2, 2, 0, 0, 1 for the 5 classes.

In conclusion, please add some suggested future works in a future works section.  Maybe look at other training methods and how that affects your GCA-CNN method.

Reviewer 2 Report

In this paper, a gated channel attention module is developed which is used to preprocess inputs to a CNN to improve the classification accuracy of SAR images. The attention module helps to implement feature fusion that helps the network to attend to the details of images. Experiments are provided to demonstrate that the method is effective and compares favorably against classic CNNs.     The paper considers a practically important problem but the novelty of this paper is minimal. CNNs have been used for classification extensively and GCA modules have been used before for various applications successfully. Merely using them in a new domain is not sufficiently novel. Moreover, experiments are limited and do not include real-world data and helpful ablative experiments. The writing also needs improvements. In conclusion, I think this paper needs significant work to make it publishable in a journal similar to Remote Sensing.

Reviewer 3 Report

This manuscript presents a new technique to improve classification accuracy using the feature fusion strategy and the Gated Channel Attention(GCA)  module. Experiments are conducted for each of these improvements by comparing them with existing methods.  However, there are a few places that can be improved to further enhance the quality of this manuscript. 

 

• The utilization of attention and gated mechanisms in remote sensing classification is not very new, please add some thoughts in the introduction or related work to help readers catch the latest development in this field, better expressing the gap this study aims to fill. 

 

• Please consider adding more details regarding the experimental setup. For example, how the training sample size is determined?  Is it based on a setup from a similar study? What is the strategy used to avoid selecting training samples from the same regions in imagery?

 

• Also, the authors mentioned that get the optimized feature map numbers, the number of iterations, and the number of layers using one benchmark imagery dataset. Does it mean these parameters can automatically be used for testing the remaining two imagery datasets? 

 

•  In terms of the experimental data presentation,  it would be more helpful if the evaluation metrics produced by the proposed model could be visualized using a confusion matrix for multi-label classification cases.

 

• Please check the terminologies used in the manuscript,  as some of them need to revise to follow the convention pattern, for example, soft-max -> softmax; sigm -> sigmod.

 

• Please rework the figures with a better and more readable layout.