Lifting Scheme-Based Deep Neural Network for Remote Sensing Scene Classification

Round 1

Reviewer 1 Report

This is a very interesting article that demonstrates, with real experiments, how the introduction of non-linear elements in convolutional networks can be an option to improve their efficiency. It is worth mentioning the detailed explanation of the most mathematical part that justifies the possibility of replacing the classical convolutional stages with non-linear stages based on lifting schemes. Highlighting the example of the application of expressions in order to follow the reasoning more clearly. It is also worth highlighting the results obtained with different image databases, with comparisons with other methods, tables and very clear and justifying graphs. However, I hope that it will continue with the future intention expressed in the conclusions given its scientific interest.

 

I think it is an excellent research article that can be published if you consider the following minor issues:

 

 

In line 169 the bibliographical reference of Wim Sweldens is missing.

 

In equations 4 and 5, I think we need to explain in more detail that each of the variables A(z), D(z), Xe(z), Xo(z), He(z), Ge(z), Ho(z) and Go(z) are. It is understood that they are the z transforms of the components of equations 2 and 3, but perhaps a brief explanation would not be too much.

 

 

I think equation 8 is not correct, the degree is k-1 not its module. Or maybe I'm not familiar with this type of nomenclature, but the usual is to use |H(z)| as its module and not as its degree.

 

 

Why in equation 12 is P(z) redefined as part of the definition of equation 5. What happens to the high pass filter part?

 

I do not understand how equation 13 is obtained. They say it is from the right side of equation 4, but it is not clear. Why Xe(z)=X(z) and Xo(z)=zX(z). I think some steps are missing.

 

In line 220, I think the meaning of the operator gcd(He(z),Ho(z)) should be better explained.

 

In figure 2, I think it would be useful to indicate that the upper branch is the branch of xe and the lower branch is the branch of xo.

 

 

In the results tables, why was the accuracy metric chosen and not other metrics? I think you should at least indicate the definition of the metric used and a brief explanation of why you chose this one and not other possible ones.

 

I think a brief explanation of what the "Delta" parameter is would not be superfluous either. It can be deduced from the tables, but it can also be added in line 329.

 

 

In the confusion matrices of figure 7 and 9, a color code has been used that does not allow us to see clearly the aspects commented in the text about erroneous values of 20% and 10%. Perhaps it would be possible to use a color map with more resolution in that critical area to better appreciate those differences. Although figures 8 and 10 help to understand the differences more clearly.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

 

The manuscript with the title “Lifting Scheme Based Deep Neural Network for Remote Sensing Image Classification” presents a study that introduces the lifting scheme into deep learning and addresses the problems that only fixed and finite wavelet bases cab replaced by the lifting scheme and the parameters can not be updated through backpropagation. The paper takes on a worthy topic and its interest in the remote sensing community. However, there is room for improvement before the manuscript goes to publication.

 

 

General comments

The manuscript, in general, needs a lot of organization. The authors spend too much effort in ththe e first part. Although it is Sections to provide a background for a new method, Sections 1 and 2 (introduction and related works) need to be summarised in one single section. There is a lot of information, but the organization is not strong and leads to some confusion.

Section 3  (methods) is focused on describing the proposed method, and later in Section 4, the experiments are introduced. I recommend revising these sections according to the journal guidelines because there is a mix of methods and results (section 4). Also, in section 4, th,ere are results that come from methods that were not described in the proper section. Datasets, study area, experiment setup need to be described in the methods section.

The analysis section (4.4) should be in a different section, perhaps as discussion, and provide elements to contrast the main highlights of your work according to what is already written in literature. As it is the analysis section is very poor and does not reflect the real contribution of your manuscript.

 

Particular

 

Figure 1 is not cited in the text. What is the purpose of the Figure?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for the Author:

 

Thank you very much for giving me the opportunity to review this manuscript entitled “Lifting Scheme Based Deep Neural Network for Remote Sensing Image Classification”. The author aims to present a new approach of introducing the nonlinearity of the lifting scheme which can enhance the computational ability of the convolution neural networks. The aim of the study is very interesting, and the manuscript is well written. However, it was difficult to find out the contribution of this method on remote sensing image analysis. As the manuscript is submitted for the Remote Sensing journal, it should be addressing readers in the field of remote sensing. Therefore, the author needs to be clearly documented the importance of this computational efficiency of the convolutional neural networks model in remote sensing. There are few major issues with the representation of the proposed work that need to be addressed and modified before the publication. The key issue I found in the manuscript is that the author provided very limited information of how does the method effect remote sensing image analysis. I therefore, believe that the issue requires a complete re-write of the manuscript and therefore my decision is to accept this manuscript with major revisions. More detailed comments are provided below:

 

Abstract:

The abstract should include the result of quantitative analysis to proof the robustness of the method. It only includes the conceptual facts. The reader also wants to know how far the method enhance the performance of convolutional neural networks.

 

Introduction:

Subsection 1.3. Contributions and Structure: Lines 104 to 118: The contribution of the lifting Scheme into convolution neural networks in the field of remote sensing is not clear. In regard to it, the contribution of the paper will be more clearer if the author explain the recent problems of using convolution neural networks on remote sensing image analysis/classification.

 

Related Works:

The author only documented the literature review on convolutional neural networks architecture. It will be necessary to review the current works of convolution neural networks applications for remote sensing image classification. Recently, many researches are contributed for remote sensing image classification. The author should mention few of those current works along with the computational challenges for high resolution aerial images.

 

Lines 154 to 157: What are those researches to improve the convolutional layers? Some examples are needed here. Moreover, the author is required to mention how does is work helps at the top of other pre-existing methods?

 

experiments: Line 200:

This section should be name as Result based on the writing style of the Remote Sensing manuscript(Please refer the section manuscript Preparation at https://www.mdpi.com/journal/remotesensing/instructions). And please make sure that if the section head starts with capital letters, then it should be maintained for all the section. Section 1, 2 and 3 heading starts with capital letter (Lines 17, 119 and 195). Under Result heading the author should make subsections: Datasets, Experiment Setup and Experiment Results.

 

Figure- 7 and Figure-8 is unreadable. It would be good if the author explain in two lines about the figures in the caption.

 

The result section only portrays the model’s performance on the dataset. But as this journal is for the remote sensing field, the readers want to see the model’s performance on the remote sensing dataset. It would be great if the author provides few examples of the classified AID images. Especially, the claims that the author had made in lines 370-375, should be followed by showing the classified images.

 

The Result section should exclusively include the results of the method on AID and CIFAR-100 data sets. The analysis should be included in the Discussion section which is missing in the manuscript.

 

The author should explain what the accuracy assessment besides the model’s accuracy on the Test set are. The results should include a detailed accuracy assessment of the classification performance on the remote sensing images and a comparison of accuracy assessment with some pre-existing models. The error matrix or Jaccard Similarity Index should be a clearer demonstration of the proposed method on the remote sensing data.

 

The Discussion section is missing in the manuscript. It should be a separate section after the Result section and include the results of the study and its importance in the regards to the previous studies and the objective of the study. The author also needs to highlight the importance of the work in a more general context and the possible limitations. The author may also highlight the future works in this section. The author can include the analysis in the Discussion section.

 

The Conclusion should be a more descriptive. The Conclusion should provide a brief overview of the manuscript and it should be self-contained so that if the reader read through the conclusion then can understand the state of the art knowledge, knowledge gaps and how the study has been formulated to overcome those gaps, how it should have helped to enhance over the previous methods and what its contribution to the remote sensing images and also future research prospects if any.

 

The major issue of this manuscript is that it shows a very limited connection to the field of remote sensing. The author need to be very careful that the method should show clearly the advantages of using convolution neural networks with Lifting scheme for remote sensing image classification. And it is also important to note down that CIFAR-100 data set is not a remote sensing data set. When the work is related to remote sensing it should either include air-borne or space-borne remotely sensed images on earth.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Many thanks to the authors for addressing all the revisions. Congratulations on your hard work.  The paper has been improved substantially. It would make a nice contribution to the RS journal.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

I appreciate your efforts to acknowledge the issues and resolve them in the modified version. The change of the manuscript title now clears the objective of the investigation to the readers. 

I would also appreciate the details included in the modified version to fill the gaps of understanding on the reader's point of view.

Considering the changes made in modified version, in my opinion the manuscript has all the attributes to be published in the Remote Sensing Journal.