Ionograms Trace Extraction Method Based on Multiscale Transformer Network

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

Thanks to the authors' response on the innovation of the paper.

In my opinion, there is moderate innovation in this paper. But considering the valuable and acceptable results that have been presented, it can be useful for readers.

Author Response

Comments 1: Thanks to the authors' response on the innovation of the paper.

In my opinion, there is moderate innovation in this paper. But considering the valuable and acceptable results that have been presented, it can be useful for readers.

Response 1: Thank you for your valuable comments, which make the content of this manuscript more complete.

The manuscript's content was supplemented according to the comments of the third reviewer. The supplementary parts have been highlighted in the manuscript.

 

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

The authors have addressed all my comments. I think the revised manuscript has improved significantly and it is suitable for publication in Remote Sensing now.

Author Response

Comments 1: The authors have addressed all my comments. I think the revised manuscript has improved significantly and it is suitable for publication in Remote Sensing now.

Response 1: Thank you for your contribution to this manuscript, which makes the content of this manuscript more comprehensive. The part marked in yellow is supplemented by referring to the opinions of the third reviewer.

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

This paper is a good study of how the different Neural Nets perform in the extraction of O and X wave height information from the ionograms. The authors present their own model IonNet that does well in most respects and better in some measures.

1) Can the authors please explain why there is a crossing point in the O and X profiles that can be seen in figure 7 (a3) and (b3)? Is this happening at an altitude when the electron cyclotron and electron plasma frequencies switch in size,i.e fpe < fce goes to fpe > fce?

2) I would like the authors to write a short section on the physics, and why the degree of accuracy is needed. To my untrained eye there isn't much difference in the physical relevance to space weather and ionospheric physics between all the different nets. One could argue that as far as the physics is concerned, all the nets are extracting almost the same information.

 

Author Response

This paper is a good study of how the different Neural Nets perform in the extraction of O and X wave height information from the ionograms. The authors present their own model IonNet that does well in most respects and better in some measures.

Comments 1: Can the authors please explain why there is a crossing point in the O and X profiles that can be seen in figure 7 (a3) and (b3)? Is this happening at an altitude when the electron cyclotron and electron plasma frequencies switch in size,i.e fpe < fce goes to fpe > fce?

Response 1: Thank you for your valuable advice. According to the existing research, researchers focus on the critical frequency of O/X wave traces. When the plasma frequency is much larger than the electron magnetic cyclotron frequency of the reflection point, the electron magnetic cyclotron frequency is twice the difference between the two critical frequencies.

In Figure 7 (a3) and (b3), the intersection points of the O/X traces are also not the true height of the reflection points. The existing model realizes the subsequent calculation of the true height of the whole electron density profile. 

This can also show that our vertical ionosonde has a good effect on ionospheric detection, the O/X wave trace is more complete, and the changes of the ionosphere can be obtained more intuitively according to the reflected echoes.

Comments 2: I would like the authors to write a short section on the physics, and why the degree of accuracy is needed. To my untrained eye there isn't much difference in the physical relevance to space weather and ionospheric physics between all the different nets. One could argue that as far as the physics is concerned, all the nets are extracting almost the same information.

Response 2: Thank you for your valuable comments. The supplementary content has been highlighted in the manuscript.

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

Comments and Suggestions for Authors

The topic of the paper is very interesting and practical. The writing structure of the paper is appropriate. The sectioning of the paper has a proper structure. The figures have sufficient resolution. The results are correctly and completely interpreted. But the most important point of this research is what is its innovation? What's new in this paper? What is the difference between this paper and previous research?

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposes a Transformer-based ionogram extraction method, and the authors validate the effectiveness of the proposed method by comparing it with FCN, U-net, and other models. I think it needs significant revisions, and my specific comments are as follows:

 

1.     In the Introduction, there are too few references related to ionogram echo extraction and the application of ionograms. The authors should reference more relevant works.

 

2. In Section 3.3, since the authors considered the class imbalance issue and chose the OhemCrossEntropy loss function, I strongly suggest adding a comparative experiment with the two different loss functions to prove the effectiveness of using the OhemCrossEntropy loss function.

 

3. Section 4.2

   - The paragraph is confusing. Firstly, are the frequency and height continuous? How did the authors come to this conclusion? In Section 4.1, the authors also mention that the detection equipment collects data at different frequency steps, so the collected frequency-height data is already a set of discrete data points. Even if understood from an image perspective, images are composed of discrete pixels, making the discrete processing here unnecessary.

   - Secondly, the authors balance data resolution and the neural network's processing capability by adjusting the frequency resolution from the original 25kHz to 50kHz, but in the previous section, the authors stated that 25kHz data was chosen for high resolution from the 25, 50, and 100kHz options. This is contradictory. Why not directly choose 50kHz data? Furthermore, in my view, a 3060 GPU with 12GB of memory should efficiently handle 25kHz resolution images. The authors should significantly revise this part.

 

4. In lines 351-358, what annotation tool was used? How long did the annotation take?

 

5. The titles of Sections 5 and 6 are redundant. Moreover, the discussion section should not simply repeat the manuscript's narrative but provide a technical or theoretical discussion of the experimental results. This includes discussing the results, the limitations, and suggestions for future research. In the conclusion section, the authors should emphasize the research contributions. These two sections need to be rewritten.

   - Additionally, in lines 585-587: “In this application, the capabilities of the multi-scale Transformer model are newly extended, leaping from processing text sequences to processing image data, especially those ionograms with high complexity and dynamic changes.” As early as 2020, the Transformer model was used to process images, and this type of model is called the Vision Transformer (ViT). The authors should not emphasize “newly extended” here.

 

6. In Line 617, “articls” -> “articles.”

Comments on the Quality of English Language

 Minor editing of English language required.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents an alternate method to extract traces from ionograms. The authors have compared their semantic approach to other neural network approaches. They have not shown why one would want to use this new method over more traditional methods like ARTIST or other image processing and heuristic approaches. they have not presented any evidence that their approach is better or worse than other methods for trace extraction rather they have concentrated on how their approach is better than alternate neural network approaches. They have not answered the important question of "what advantages does this method offer??" This is the big question, not is it better or worse than other CNN and NLP approaches.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The English is adequate.