Boosting Semantic Segmentation of Remote Sensing Images by Introducing Edge Extraction Network and Spectral Indices

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors introduce edge information into the deep learing algorithm to semanticly segment the remote sensing imagery. Several comments are given as follows.

1. The parametrs show in tables did not give any explanation. It is hard to understand that the talbes want to show.

2. In Figure 6 and 7, the x axises did not have any explanation.

3. Though different deep learing algorithms were used to compare the accuracies, the performace evaluations are supposed to be given.

4. In Figure 6 and 7, the proposed apparoch seems to have a better peroframce on a particular image but for other images, the performances are not obvious. Any explanation?

5. The authors did a lot of work but the processed results are supposed to be reorganized.

Author Response

We thank the reviewers for their comments on our manuscript. We have done our best to improve the manuscript and hope that the corrected manuscript will be recognized. The response to the reviewers' comments can be found in Response 1.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes a holistic attention edge detection network by augmenting the network input with multiple nonlinear spectral indices (including vegetation and water indices), called HAE-RNet.

My concerns and suggestions are listed as follows:

1.       The authors claimed that DSM could be used in the proposed network, but it is a bit unclear how DSM is used. I do not find DSM in the architecture of the proposed network as shown in Figure 2. Is DSM (or a depth map) straightforwardly used as one of the inputs to the proposed network?

2.       It is better to review some methods that achieve SOTA performances on the Vaihingen datasets, such as [a,b]. And if possible, it is better to make comparisons with them.

[a]SCE-Net: Self- and Cross-Enhancement Network for Single-View Height Estimation and Semantic Segmentation. Remote Sensing 2022.

[b]MCFINet: Multi-Depth Convolution Network with Shallow-Deep Feature Integration for Semantic Labeling in Remote Sensing Images, IEEE Geoscience and Remote Sensing Letters 2022.

3.       Why is the result with “9c” channels is worse than that with “6c-2” channels in Table 2? Intuitively, the result with “9c” channels should be better or close to that with “6c-2” channels because “9c” channels contain more information?

4.       The authors are suggested to show some intermediate visualization results of the proposed method on edge detection, for demonstrating the effectiveness of the introduced holistic attention edge detection block.

5.       Section 4 could be combined with Section 5 together.

Comments on the Quality of English Language

The English writing should be polished further.

Author Response

We thank the reviewers for their comments on our manuscript. We have done our best to improve the manuscript and hope that the corrected manuscript will be recognized. The response to the reviewers' comments can be found in Response 2.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors  

This paper enhances the semantic segmentation performance of remote sensing images by augmenting the network input with multiple nonlinear spectral indices and introducing a novel global edge detection network, HAE-RNet. Experimental results show that the NIR-NDWI/DSM-GNDVI-R-G-B (6C-2) band combination produces the best segmentation results for both the GID and Vaihingen datasets. However, there are several issues that need to be addressed:

1. The term "Remotely Sensed Imagery" is not commonly used and needs to be modified.
2. Figure 1 is unclear and the black outlines are not visible enough. It needs to be more clearly expressed.
3. The contribution of the paper needs to be expressed more clearly in the introduction, and the design intention needs to be expressed at a higher level.
4. The innovative aspects of the paper need to be listed in the introduction.
5. The review of related research needs to be more in-depth and up-to-date, including the latest research results such as "Building extraction from remote sensing images with sparse token transformers" and "Efficient transformer for remote sensing image segmentation" based on Transformers and "RSprompter" based on large models.
6. The similarities and differences between this method and other research need to be highlighted in the description of related research, as well as the original intention of this method design.
7. The differences between this method and UNet need to be emphasized in the method section.
8. It is unclear from Figure 2 whether two different sets of images enter different backbones.
9. The description of edge-related methods in the method section is too brief and needs to highlight the main innovations.
10. If the edge serves as the output, it needs to be expressed in the method figure.
11. More comparisons with current advanced methods need to be made.

Comments on the Quality of English Language

N/A

Author Response

We thank the reviewers for their comments on our manuscript. We have done our best to improve the manuscript and hope that the corrected manuscript will be recognized. The response to the reviewers' comments can be found in Response 3.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

A good approach to using CNN in remote sensing applications. The authors are to be congratulated.

Author Response

We thank the reviewers for their comments on our manuscript. We have done our best to improve the manuscript and hope that the corrected manuscript will be recognized. The response to the reviewers' comments can be found in Response 4.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised version did meet the reviewer's opinions. No more comments are against the paper.

Author Response

Thank you very much for recognising our revision work. In the future, we will continue to work hard to achieve more results and thank you again for recognising our work.

Reviewer 2 Report

Comments and Suggestions for Authors

Most of my concerns have been addressed.

In the current version, the journal names of many references in the reference list seem to be wrong, e.g., [14], [15], [42], etc.  The authors are suggested to check all the references carefully.

Comments on the Quality of English Language

The authors are suggested to polish the writing thoroughly, and revise the reference list carefully.

Author Response

We thank the reviewers for their comments on our manuscript. We have tried our best to improve the manuscript and hope that the corrected manuscript will be recognised and the response to the reviewers' comments is in Response 2.

Author Response File: Author Response.pdf