Global–Local Information Fusion Network for Road Extraction: Bridging the Gap in Accurate Road Segmentation in China

Round 1

Reviewer 1 Report (Previous Reviewer 3)

Line449-451: Why did you choose these models to compare with the models in this paper? Some of the models used for comparison are presented here for the first time.

 

The Introduction section is too long, not logical enough and a bit redundant, consider adding subheadings to improve readability and logic.

 

Discussion sections can be added to the text.

 

In addition to the improvement in accuracy, has the model presented in this paper improved in speed?

 

In summary, I was happy to review your manuscript: “Global-Local Information Fusion Network for Road Extraction: Bridging the Gap in Accurate Road Segmentation in China”. This one is a big improvement on the last one, with detailed answers to each of the comments and refinements to the article, but there are still some minor problems, which are expected to be improved.  Accept after minor revision (corrections to minor methodological errors and text editing)

 Moderate editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

1. What is the exact difference between road extraction task for China compared that of other countries?

2.  In P3, r72, "Furthermore, the generalization capability of CNN-based models and the potential loss of road feature information due to pooling layers are ongoing challenges that must be addressed." Is there any reference to support this claim?

3.In figure 4, it is hard to distinguish roads from human eyes. Therefore, it is suggested to add the according groundtruth for better demonstration.

4. In figure 5, there are no "GIE" or "LIE" marked, while these two are the key modules in this method.

5. The running time is suggested to be compared among all the presented methods and it is also a key criteria to evaluate the efficency of road extraction methods.

1. P3 R72, "model's ability" should be "the ability of models"

2. P6 R196, the characteristics of CHN6-CUG dataset should not be "limited data volume" or "lack of public datasets", they are the shortage of other existing datasets.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Introduction：

The introduction is best to cite and introduce the CNN+transformer methods, as well as the traditional LBP methods.

The references are not rich enough, and the following references are recommended.

(1) G. Xu, T. Song, X. Sun and C. Gao, "TransMIN: Transformer-Guided Multi-Interaction Network for Remote Sensing Object Detection," in IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1-5, 2023, Art no. 6000505, doi: 10.1109/LGRS.2022.3230973.

(2) Automatic extraction of urban road boundaries using diverse LBP features. National Remote Sensing Bulletin, 26(3): 541-554 DOI: 10.11834/jrs.20209228.

Method：

In formula (6)   needs to be explained.

Experiments：

Use more recent and advanced baseline methods for comparison.

The ablation study is limited to evaluating the MSF and SC-Att modules. Additional experiments could help analyze the contribution of the other key components proposed.

Better report and compare the computational complexity.

Conclusion：

The conclusion effectively summarizes the key findings and contributions of the paper. To further enhance the conclusion, consider the following points:

1. Discuss the broader implications and potential applications of the research findings in the context of road traffic, humanitarian rescue, and environmental monitoring.

2. Identify any potential avenues for future research or extensions to the GLNet model, such as exploring its applicability to different geographical regions or investigating its performance on other types of road networks.

Can be improved. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

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 for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments on reviewing the following manuscript for Remote Sensing

Global-Local Information Fusion Network for Road Extraction: Bridging the Gap in Accurate Road Segmentation in China

 

This manuscript proposes a new method for road extraction. The topic is interesting, but I recommend the rejection of this paper as it is flawed in various aspects. My comments are as follows:

 

1. The first sentence of the abstract can be deleted.

 

2. The full names of terms like GIE and LIE should also be written in the Introduction.

 

3. The authors present the Introduction where related work is summarized. However, many citations are about semantic segmentation or deep learning, and there is an insufficient summary of the current research status of road extraction. Therefore, the insufficient extraction of road edge details may have been solved.

 

4. This article evaluates the performance of GLNet by comparing it with other methods, including DANet, Deeplabv3+, PSPNet, Segformer-b5, and UNet. However, these are not specific methods for road extraction. In recent years, many deep learning methods specifically for road extraction have been published, such as (1) and (2). The GLNet method should be compared with these algorithms.

 

(1) Zhu, Q.; Zhang, Y.; Wang, L.; Zhong, Y.; Li, D. A Global Context-aware and Batch-independent Network for road extraction from VHR satellite imagery. ISPRS J. Photogramm. Remote Sens. 2021, 175, 353-365.

(2) Li, X.; Wang, Y.; Zhang, L.; Liu, S.; Li, Y. Topology-Enhanced Urban Road Extraction via a Geographic Feature-Enhanced Network. IEEE Trans. Geosci. Remote Sens. 2020, PP, 1-12.

 

5. What is the basis for the experimental setup in Section 4.1? If the parameters are changed, will other comparison methods be better than the method proposed in this paper?

 

6. English is also needed to improve. Lines 131-136 are with the same subject, hence should be merged otherwise different subjects are needed!

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Line155 and line 501: The text is not aligned.

 

Line311-313: you mention F1, F2 and F3, but in Figure 9 you can only see input F1 and output F3, not F2, which should be noted in the figure 9.

Line311:Wrong letter in line 331.

Line333-334: Road extracting from images can be handled as a binary classification problem, so it is more appropriate to use a cross-entropy loss function. You should state why when choosing a loss function.

 

Line355-356: You say in order to prevent overfitting and enhance the robustness of the models, various data augmentation techniques such as random scaling, random flipping, and random cropping were employed during the training process. Here, you should detail what the specific parameters are in each step of the data enhancement, and in what form and with what parameters the final picture is entered into the model.

 

Line358-359: You said road extraction from images can be effectively treated as a binary segmentation problem, why?

 

Line384-389: You are all about Intersection over Union (IoU), but there is an error in writing the calculation equation (14), it is not IoU but OA.

 

In the Evaluation Metrics section, I think evaluation metrics such as ROC curve and AUC can also be introduced, and the ROC curve as an image will be more intuitive to reflect the performance of each model.

 

Line448-449: You say that it is challenging to conduct experiments on the RDCME dataset due to low resolution and insufficient data compared to the CHN6-CUG dataset, but the experimental results show that the model outperforms the CHN6-CUG dataset on the RDCME dataset, why?

 

Line495-498: You said the performance of the model is slightly lower than that on DANet in terms of OA may be attributed to the simplicity of the IF module, which leads to the loss of crucial road details during the fusion process. Why?

 

Others:

1、The challenges outlined in the introduction section and the contributions in this article should ensure logical consistency and relevance between them.

2、Enumerating the challenges faced by shadowing road features is not addressed in the paper.

3、The speed of improvement mentioned in the paper is more convincing in the comparison of experimental results.

 

In summary, I was happy to review your manuscript: “Global-Local Information Fusion Network for Road Extraction: Bridging the Gap in Accurate Road Segmentation in China”. Overall, it was clear that a lot of work has gone into this research, such as described in detail the various modules and functions of the model, and compared the experimental results of the GLNet model with several models. owever, the manuscript still need to be improved. Recommendation is Reconsider after major revision.

Moderate editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.docx