A Bidirectional Deep-Learning-Based Spectral Attention Mechanism for Hyperspectral Data Classification
Round 1
Reviewer 1 Report
Dear authors,
After reviewing your manuscript I found some things to need a clearer description or explantion, such as:
- What advantage do you find mixing PCA with DL? because DL can deal fine with huge datasets and multiple features.
- How do you proof that your method acts as an efficient noise reduction component? because some noise could be important features not recognized as that.
- Why do you say the overall execution times are comparable? while table 6 shows your method is almost 3 times slower than other methods.
- Also, I would like to know what is the first % training data in figure 5? because you mention 10% throught the whole manuscript.
It is all for now,
Best regards,
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This study proposes a deep-learning-based bidirectional spectral attention and classification method for hyperspectral images. The expression and structure of the manuscript are relatively acceptable, but the experiment and analysis parts need to be strengthened. The following questions need to be solved before it goes further.
- Please reduce the number of keywords.
- Line145: Why 3×3×30?
- Line148: What does K mean?
- Lines 268-272: there is too little discussion on the results of different methods, and the advantages of the methods are not presented and analyzed.
- Lines 289-291: the experiments in this manuscript did not prove the importance of PCA. At least the comparative experiments of the proposed model under the premise of PCA and without PCA were not done.
- Class specific accuracies of Salinas dataset is lost.
- 4.2. Parameter Tuning and Experimental Setup: the proposed method has several parameters that need to be set. In this study, these parameters are selected based on experience. However, the impact of these parameters on the classification results is not discussed. I think these are of interest to readers. In fact, the accuracy of the proposed method is not much improved compared to other methods, so I am worried that changes in these parameters will have a greater impact on the results, thereby affecting the practical applications of the method.
- Comparative experiments are not sufficient, and some of the latest advanced hyperspectral image classification methods need to be considered.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Generally, this manuscript is suitable for acceptance in remote sensing journal. Because authors proposed a new model by using 3D Conv and bi-directional LSTM, and their reason why used those layers is also reasonable.
Maybe in the proposed model part, authors need to explain more structurally about their model. Because, in my opinion, I miss some detailed explanations when I read Figure 1 and the explanation of the proposed model.
In addition, some minor problems should be addressed:
• The introduction part was well organized, but I didn't find any information about the advantages of hyperspectral images for land cover classification. Because this manuscript is about hyperspectral data for land cover classification, the information about the advantages of hyperspectral data for land cover classification is better to explain.
• Line 143. The authors need to add a legend in Figure 1 to explain some symbols, like X circled, and cross circled.
• In the proposed model, after the extracted features are fed into Bi-directional LSTM, the authors use SoftMax activation function, what is the reason the authors use SoftMax activation function at that position? How about the final FNN (C nodes), did the authors also use SoftMax as the final activation for classification?
• In Table 6 the authors just show the overall execution time. The information regarding time per-epoch and the total number of parameters for each model needs to be added because those are also important information for the comparison analysis to know the effectiveness and efficiency of the proposed model.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Dear authors,
After reading your responses to my inquieries, I felt satisfyed about them. You have improve the manuscript's soundness and content, as it was my wish.
Congratulations to you,
Best regards,
Author Response
We are very grateful for all you comments on our manuscript. Upon making the necessary changes advised by you, our manuscript looks to be in a better shape for publication.
We once again whole heartedly thank you for working with us and helping us get our manuscript a step closer towards publication.
Reviewer 2 Report
Thank you very much for the replies to my concerns. However, there are still several problems to be discussed with the authors.
- Lines 309-314: The comparison of the classification results of different methods in the manuscript is only described in one sentence, which is unreasonable. The authors need to clearly point out and analyze the advantages of this method, not just list figures and tables with weak analysis.
- Lines 337-355: Please add references, at least two of which you have published before.
- The proposed method has several parameters to be set. The authors describe that all parameters have been optimally selected. Does the author ensure that the parameters of the comparison method have also been optimally selected? In fact, compared with other methods, the accuracy of this method has not been greatly improved. Thus, the influence of these parameters on the classification results needs to be discussed. On the one hand, it proves that the method is indeed effective; on the other hand, it is a reference for readers.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
