Change Detection for Heterogeneous Remote Sensing Images with Improved Training of Hierarchical Extreme Learning Machine (HELM)

Round 1

Reviewer 1 Report

Dear Authors,

I have reviewed the paper entitled "Change detection for heterogeneous remote sensing images with improved training of hierarchical extreme learning machine (HELM)" submitted to the Remote Sensing Journal. It is an interesting paper which deals with a new change detection framework. The quality of the work is almost good. However, some aspects of the methodology and case study description could be reviewed. Below are some comments:

• My main concerns are related to the description of the methodology part. It should be prepared with sufficient detail to allow others to replicate and build on published results, while in the submitted paper, there is no information about training and testing parameters.
• I leave it to the authors to consider whether it would be a good practice to make code, etc., available online (as a supplementary file or using Github. Such approach would strengthen the scientific quality and allow for scientific discussion with readers and researchers. Also, It could fulfill the description of used parameters and make it possible to reproduce results for further comparisons.
• The readability of figure 3 is relatively low. Authors should think about choosing different colors or changing the size of the text.
• Fig 10 ->Because of low visual differences between g) and h), it would be good if both images were colored. Also, the reference image could be colored.
• Section 2.2 -> The equations have not been used for the first time. The reference should be added.
• Please check for the additional spaces in the text, e.g., line 294.

 

Sincerely,

Reviewer

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Here are some of the issues that the authors need to address to improve the quality of the paper.
1. The algorithm presented has not any novelty.
2. The authors are suggested to measure the performance through the analytical model and simulation to verify the correctness and usefulness of their analytical model.
3. The 'conclusions' is a key component of the paper. It should complement the 'abstract' and normally used by experts to value the paper's engineering content. In general, it should sum up the most important outcomes of the paper. In addition, some limitations of this study and the future works should be put on this section.
4. The manuscript is not well organized. The introduction section must introduce the status and motivation of this work, and summarize with a paragraph about this paper.
5. For better readability and comprehensibility, please highlight the key research problem of this work clearly.
6. The reviewer thus will not recommend the work to be published at this time.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, the authors developed a remote sensing image change detection method based on hierarchical ELM. I am afraid this paper can not be published in current form. 

1. The novelty of this paper is quite limited. HELM has been developed and applied to image analysis. This paper only applies this method to a specific application. 

2. Most recent neural-network-based remote sensing image change detection methods are not introduced in the introduction. More specifically, recently a supervised PCA-Net was proposed for SAR image change detection, which was not included in the introduction. Therefore, the authors should introduce most recent work in the related works. 

3. In the section 3, the authors introduce more details on sample selection. It is not sure the motivation of the sample selection by clustering. Currently, there are a number of labeled data for supervised learning, which obtains promising result. Can this model be trained by supervised manner? If can, can the author show the results?

4. In the pannel (f) of Figure 10-13, the authors highlight the class of changing by different colours. Why do the authors only highlight the changing class of the proposed method? How does the proposed method output the changing class? Please the authors explain more details. 

5. Please compare the proposed method with recently proposed methods.

6. Do the authors plan to release the code and dataset if the paper is published? 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Authors,
I have rereviewed the paper entitled "Change detection for heterogeneous remote sensing images with improved training of hierarchical extreme learning machine (HELM)" submitted to the Remote Sensing Journal.  All my previous comments have been addressed. All my concerns were clarified.

Sincerely,
Reviewer

 

Author Response

Dear Reviewer：

Thank you again for giving your valuable time to review my manuscript!

Kind regards.

Your sincerely,

Yuqi Tang

Reviewer 2 Report

The manuscript is well organized in its current form, I suggest to accept it.

Author Response

Dear Reviewer：

Thank you again for giving your valuable time to review my manuscript!

Kind regards.

Your sincerely,

Yuqi Tang

Reviewer 3 Report

In this paper, the authors made a serious revision, but they may not solve all my concerns. 

1. The novelty of the proposed method is still quite limited. The author only applied HELM to a specific task.
2. The authors compared the supervised method with the proposed unsupervised method. But the results can not convince me, since the details on the supervised method is unknown, including how to select training samples. If the samples are well selected for training, the results on supervised learning may not be worse than the unsupervised method. 
3. The proposed method is an unsupervised method. How can the proposed method predict the label of land-covers? In fact, an unsupervised method can only predict a sample belonging to which class, but can not predict the semantic label. Because there is no semantic information is introduced in the training stage. 
4. The authors mentioned that they can not release the code even the paper is published since they still have some relevant experiments going on at the same time. Can I understand that the experiments related to the proposed method are not completed? Otherwise, why not release the code if the paper is published?

Hopefully, the authors can solve my concerns in the next round of revision. 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 3

Reviewer 3 Report

In this paper, the authors made a revision and respond to my concerns. However, there are two concerns that the author did not solve well. 

1. The novelty of this paper is quite limited. The author did not improve the method.
2. The author did not show the comparison results with the supervised methods in the revised paper. Furthermore, the author's response on the results of supervised methods can not convince me. Actually, there have been many supversied methods proposed for SAR image change detection. Most of them can obtain better performance on SAR image change detection than the unsupervised method. But the author declared in the response that the supervised methods will not be worse than the unsupervised methods if well-selected training samples are used. I can't agree with this opinion. 
3. I do suggest the authors show the comparison results with the supervised change detection methods, e.g. supervised PCA-Net, and other supervised CNN methods. If the author can not give promising comparison results, I have to reject this paper. 
4. Please the author prepare for the link to the code and data if they want to release the code and data.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf