Semantic Segmentation of China’s Coastal Wetlands Based on Sentinel-2 and Segformer

Round 1

Reviewer 1 Report

I have reviewed the manuscript titled " Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer". This manuscript discusses the application of the SegFormer model and Sentinel-2 data for the classification and segmentation of coastal wetlands in Yancheng, Jiangsu, China. The study aims to extract wetland information efficiently and accurately, which is crucial for wetland ecosystem research and management.

I have the following observations on this MS. 

The MS does contribute new in terms of methodology - a set of well-known methods have been applied for wetland change scenarios in terms of land and deriving forces and these methods are important as well.

The introduction is weak, and the method section is trivial and vague in places. More recent literature work is required.

I see a fruitful discussion on the generated datasets. But the introduction must be improved and the scientific problem has to be clearly identified and addressed.

I do see little novelty in both scientific findings or methodological improvement. The authors clearly state the scientific significance of mapping wetland and explain the relationship of economics, rather than saying something very broad.

The authors should explain more about the classification procedure adopted by the source. For example, whether classifiers were trained separately for different images, or a universal classifier was trained and applied for all years.

The authors did not provide the overall classification accuracy also there was no accuracy assessment of the built-up and water type and subsequent change detection. Without accuracy, it may be difficult to guarantee the correctness of the final conclusion. I recommend that the authors provide classification accuracy for different land types, as well as change detection accuracy.

 

In Discussion, "Authors should discuss the results and how they can be interpreted in the perspective of previous studies and the working hypotheses.

Author Response

 

Manuscript ID: remotesensing-2451164

Paper Title: Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer

Authors: XuFeng Lin , YouWei Cheng , Gong Chen , WenJing Chen , Rong Chen , DeMin Gao , YinLong Zhang , YongBo Wu *

 

 

Dear Editors/Reviewers:

We would like to take this opportunity to thank you for your precious time and great efforts in handling the above referenced manuscript. Thanks also go to the anonymous reviewers for their valuable comments that helped to improve the quality of the paper. Overall the comments have been fair, encouraging and constructive. We have learned much from it. We have thought over the comments and revised the paper carefully.

 

Enclosed please find a summary of changes, a list of one-to-one responses to the comments of reviewers and the editor. Please kindly let us know if there are any additional comments/concerns. We are looking forward to hearing from you.

 

 

 

Sincerely,

Xufeng Lin, Zhongyuan Li, Wenjing Chen and Demin Gao  

July 12, 2023

 

 

 

Response to Reviewer ’s comments

 

1. The introduction is weak, and the method section is trivial and vague in places. More recent literature work is required..

 

Our response:

<> Thank you for the detailed review. We have revised the introduction and methods, and introduced more latest related literature research. Some of the documents are as follows.

Li, X.; Xu, F.; Liu, F.; Lyu, X.; Tong, Y.; Xu, Z.; Zhou, J. A Synergistical Attention Model for Semantic Segmentation of Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing 2023, 61, 1-16.

He, X.; Zhou, Y.; Zhao, J.; Zhang, D.; Yao, R.; Xue, Y. Swin Transformer Embedding UNet for Remote Sensing Image Semantic Segmentation. IEEE Transactions on Geoscience and Remote Sensing 2022, 60, 1-15.

Raza, A.; Huo, H.; Fang, T. EUNet-CD: Efficient UNet++ for Change Detection of Very High-Resolution Remote Sensing Images. IEEE Geoscience and Remote Sensing Letters 2022, 19, 1-5.

 

2. I see a fruitful discussion on the generated datasets. But the introduction must be improved and the scientific problem has to be clearly identified and addressed..

 

Our response:

<> Thank you for the detailed review. We have revised the introduction, and more clearly expounded the research significance of coastal wetlands and the reasons for related ecological values. Introduction begins with the background, and then expounds the demand related to the value of ecological services and the development process of relevant research. From relatively traditional statistical methods to machine learning, and then to deep learning methods in recent years.

 

3. I do see little novelty in both scientific findings or methodological improvement. The authors clearly state the scientific significance of mapping wetland and explain the relationship of economics, rather than saying something very broad..

 

Our response:

<> Thank you very much for your advice. Our research combines the actual situation to study the coastal wetland, expounds the existing models and methods in application as a whole, and discusses some key problems. We have revised the contents of the article, more specifically expounded the scientific significance of wetland mapping, and further expounded the economic relationship. The evaluation standard of wetland system service value is a challenging and in-depth research topic, so this paper has no progress explanation. The contents of some modifications are as follows.

The assessment of ecosystem service value serves as a pivotal bridge connecting ecosystem research and management decision-making. Its primary objective is to comprehensively comprehend the current status and dynamic trends of ecosystem services. The creation of a wetland map plays a crucial role in establishing the indis-pensable spatial and geographical foundation for the quantitative evaluation and cal-culation of wetland ecological service value. It also facilitates a comprehensive under-standing of the position and function of coastal wetlands within the broader ecosystem, encompassing factors such as hydrological characteristics and species diversi-ty.Furthermore, the generation of a wetland map provides fundamental data and in-formation support for the delineation of wetland protection zones and the implemen-tation of wetland protection and restoration measures. This invaluable resource ena-bles decision-makers to holistically consider the ecological value of wetlands and averts the destruction and wastage of valuable wetland resources.

The value of wetland ecosystem services serves as a monetized representation illustrating the collective contribution of wetland ecosys-tems to human well-being. It encompasses an array of services such as product supply, ecological regulation, and spiritual and cultural benefits. When evaluating the wetland ecosystem service value of Yancheng, it is crucial to select appropriate evaluation in-dices and systems aligned with the region's available resources and market conditions. The wetland mapping provides valuable spatial and geographic foundations that fa-cilitate effective assessment of the wetland ecosystem service value.

Regarding the significant components of Yancheng's coastal wetland, namely Aquaculture, farmland, and mudflat, their construction should prioritize enhancing the ecosystem service value concerning product supply, water flow regulation, and water purification. As for the special vegetation occupying a comparatively smaller proportion, a comprehensive assessment of their ecosystem service value should con-centrate on aspects related to species conservation, tourism and recreation, as well as scientific and educational services..

 

 

4. The authors should explain more about the classification procedure adopted by the source. For example, whether classifiers were trained separately for different images, or a universal classifier was trained and applied for all years.

 

Our response:

<> Thank you for the detailed review. Our data involves years of information. Therefore, the training model is universal. We have modified the introduction, made a preliminary introduction to various classifiers, and introduced the latest findings of the model in this field.

 

5. The authors did not provide the overall classification accuracy also there was no accuracy assessment of the built-up and water type and subsequent change detection. Without accuracy, it may be difficult to guarantee the correctness of the final conclusion. I recommend that the authors provide classification accuracy for different land types, as well as change detection accuracy.

 

Our response:

<> I believe that we didn't give you a good description, so you were confused. We used mIou,mAcc, m-Fscore and other indicators to show the overall classification accuracy. Our original intention here is to discuss the land types that are difficult to divide and put forward our solutions, and verify them through the results. For the land types with high classification accuracy, the change of accuracy is not particularly great. For the evaluation of ecosystem value, these land types with high value and difficult to divide are more critical issues. We explained the change of precision in the text at the top of the table, so it was not shown in the table. For the sake of aesthetics and the display of the segmentation effect for difficult cases, we have not shown the changes of other land types in the table.

 

6. In Discussion, "Authors should discuss the results and how they can be interpreted in the perspective of previous studies and the working hypotheses..

 

Our response:

<> Thank you for the detailed review. We have further explained the various performances and results. The specific paragraphs are as follows.

In this study, we utilized high-resolution images provided by Sentinel-2 to count the percentage distribution of feature classes in the study area, enabling subsequent calculation of ecological values. The specific proportional information is illustrated in Figure 13.

Figure 13 reveals that Aquaculture, Farmland, and Tidal flat constitute the pre-dominant components of Yancheng's coastal wetland, accounting for 42.2%, 29.7%, and 14.7% respectively. Notably, S. salsa and P. australis, which represent unique veg-etation within the wetland, encompass just 1% and 3.6% respectively, totaling an area of 44.79 km2 and 100.84 km2. The utilization of deep learning techniques enables the swift detection of changes in Yancheng's coastal wetland, as well as variations in the proportion of different species. The value of wetland ecosystem services serves as a monetized representation illustrating the collective contribution of wetland ecosys-tems to human well-being. It encompasses an array of services such as product supply, ecological regulation, and spiritual and cultural benefits. When evaluating the wetland ecosystem service value of Yancheng, it is crucial to select appropriate evaluation in-dices and systems aligned with the region's available resources and market conditions. The wetland mapping provides valuable spatial and geographic foundations that fa-cilitate effective assessment of the wetland ecosystem service value.

Regarding the significant components of Yancheng's coastal wetland, namely Aquaculture, farmland, and mudflat, their construction should prioritize enhancing the ecosystem service value concerning product supply, water flow regulation, and water purification. As for the special vegetation occupying a comparatively smaller proportion, a comprehensive assessment of their ecosystem service value should con-centrate on aspects related to species conservation, tourism and recreation, as well as scientific and educational services..

 

 

Author Response File: Author Response.doc

Reviewer 2 Report

Efficient extraction of wetland information is crucial for understanding and managing the ever-changing wetland environment. In this paper, evaluation indicators were used to assess the performance of the SegFormer deep learning model, used for the coastal wetlands in Yancheng, China. It is demonstrated that the model outperformed existing models in terms of accurately extracting small-scale features. The study addresses the challenge of imbalanced wetland categories by combining different loss functions to improve classification accuracy. The results provide valuable insights for wetland research and offer technical support for segmenting coastal wetlands using deep learning techniques.  

  The paper studies a practically important problem and the experiments are extensive and convincing. I have the following comments to be addressed for the next round of reviews.   1. UNet-based architectures have been used extensively for semantic segmentation, especially in medical fields. For example:   a. After-unet: Axial fusion transformer unet for medical image segmentation. CVPR 2023   b. Domain Adaptation for the Segmentation of Confidential Medical Images, BMVC 2023   c. Mixed transformer u-net for medical image segmentation. ICASSP 2022   d. ConvUNeXt: An efficient convolution neural network for medical image segmentation. Knowledge-Based Systems, 253, p.109512.   e. U-Net-Based medical image segmentation. Journal of Healthcare Engineering, 2022.   I think the above works can be discussed in the introduction section to give a broader perspective to readers about potential of UNet which has been used in the experiments of the paper   2. Please do a pass and improve the quality of the figures. I think they are very helpful and having high-quality versions is going to be helpful.   3. In Figures that you visualize segmentation results, is it possible to add a mean IoU in at the bottom of each image to make a quantitative comparison about the segmentation quality possible?   4. Please run your code several times and report both the average performance and the standard deviation on your tables to make the comparison statistically meaningful.      5. Please arrange for releasing the code on a public domain such as GitHub so other researchers can reproduce the results conveniently for future research.   

Author Response

 

Manuscript ID: remotesensing-2451164

Paper Title: Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer

Authors: XuFeng Lin , YouWei Cheng , Gong Chen , WenJing Chen , Rong Chen , DeMin Gao , YinLong Zhang , YongBo Wu *

 

 

Dear Editors/Reviewers:

We would like to take this opportunity to thank you for your precious time and great efforts in handling the above referenced manuscript. Thanks also go to the anonymous reviewers for their valuable comments that helped to improve the quality of the paper. Overall the comments have been fair, encouraging and constructive. We have learned much from it. We have thought over the comments and revised the paper carefully.

 

Enclosed please find a summary of changes, a list of one-to-one responses to the comments of reviewers and the editor. Please kindly let us know if there are any additional comments/concerns. We are looking forward to hearing from you.

 

 

 

Sincerely,

Xufeng Lin, Zhongyuan Li, Wenjing Chen and Demin Gao  

July 12, 2023

 

 

 

Response to Reviewer ’s comments

 

1. UNet-based architectures have been used extensively for semantic segmentation, especially in medical fields. For example:   a. After-unet: Axial fusion transformer unet for medical image segmentation. CVPR 2023   b. Domain Adaptation for the Segmentation of Confidential Medical Images, BMVC 2023   c. Mixed transformer u-net for medical image segmentation. ICASSP 2022   d. ConvUNeXt: An efficient convolution neural network for medical image segmentation. Knowledge-Based Systems, 253, p.109512.   e. U-Net-Based medical image segmentation. Journal of Healthcare Engineering, 2022.   I think the above works can be discussed in the introduction section to give a broader perspective to readers about potential of UNet which has been used in the experiments of the paper  

 

Our response:

<> Thank you for the detailed review. We have modified the contents of the introduction and added some models about Unet architecture in the field of remote sensing. Some cited documents are as follows.

He, X.; Zhou, Y.; Zhao, J.; Zhang, D.; Yao, R.; Xue, Y. Swin Transformer Embedding UNet for Remote Sensing Image Semantic Segmentation. IEEE Transactions on Geoscience and Remote Sensing 2022, 60, 1-15.

 Lv, Z.; Huang, H.; Gao, L.; Benediktsson, J.A.; Zhao, M.; Shi, C. Simple Multiscale UNet for Change Detection With Heterogeneous Remote Sensing Images. IEEE Geoscience and Remote Sensing Letters 2022, 19, 1-5.

Raza, A.; Huo, H.; Fang, T. EUNet-CD: Efficient UNet++ for Change Detection of Very High-Resolution Remote Sensing Images. IEEE Geoscience and Remote Sensing Letters 2022, 19, 1-5.

 

2. Please do a pass and improve the quality of the figures. I think they are very helpful and having high-quality versions is going to be helpful. 

 

Our response:

<> Thank you for the detailed review. We have adjusted the resolution of the picture to improve its quality.

 

3. In Figures that you visualize segmentation results, is it possible to add a mean IoU in at the bottom of each image to make a quantitative comparison about the segmentation quality possible? 

 

Our response:

<> Thank you for the detailed review. I have added blank lines at the back of two tables. We have added numbers to the picture to better illustrate IoU.

 

4. Please run your code several times and report both the average performance and the standard deviation on your tables to make the comparison statistically meaningful.

 

Our response:

<> Thank you for the detailed review.These results are obtained after many times of execution, and the results of running the code each time are basically the same, and the error can be reduced to three decimal places, so the results can be basically considered to be the same. I believe that some models such as unet confuse you. We can't deny the good effect of this model in related fields, but these results are based on the same experimental parameters. Under the same experimental conditions, the performance of unet is indeed a little worse, but this does not explain the advantages and disadvantages of this model, only that it is not recommended first in the segmentation of coastal wetlands. Because deep learning is not well interpretable, there are many reasons for this. Different servers and environments may bring different results. I hope you can understand.

 

5. Please arrange for releasing the code on a public domain such as GitHub so other researchers can reproduce the results conveniently for future research.

 

Our response:

<> Thank you for your comments. We will sort out the code and submit it in the following work.

 

 

Author Response File: Author Response.doc

Reviewer 3 Report

In this paper, a SegFormer-based information extraction model of Yancheng Binhai wetland is constructed using the high-resolution remote sensing images provided by Sentinel-2 satellite. This article presents a well-organized, distinctly defined, and exhaustive study on the methodology. However, certain issues require attention.

1.  The visual representation of Table 3 is unappealing, and the removal of the first row may be beneficial.
2. The predicted image sizes of various models in Figure 11 are diminutive and uniform, warranting the highlighting of distinct segments of the models to enhance the readers' experience. 
3. The bolding of the corresponding indicators in Table 7 should align with Table 6. 
4. Furthermore, the article should furnish contextual information on the specific challenges encountered by the wetland ecosystem in Yancheng, Jiangsu.

Minor editing of English language required.

Author Response

 

Manuscript ID: remotesensing-2451164

Paper Title: Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer

Authors: XuFeng Lin , YouWei Cheng , Gong Chen , WenJing Chen , Rong Chen , DeMin Gao , YinLong Zhang , YongBo Wu *

 

 

Dear Editors/Reviewers:

We would like to take this opportunity to thank you for your precious time and great efforts in handling the above referenced manuscript. Thanks also go to the anonymous reviewers for their valuable comments that helped to improve the quality of the paper. Overall the comments have been fair, encouraging and constructive. We have learned much from it. We have thought over the comments and revised the paper carefully.

 

Enclosed please find a summary of changes, a list of one-to-one responses to the comments of reviewers and the editor. Please kindly let us know if there are any additional comments/concerns. We are looking forward to hearing from you.

 

 

 

Sincerely,

Xufeng Lin, Zhongyuan Li, Wenjing Chen and Demin Gao  

July 12, 2023

 

 

 

Response to Reviewer ’s comments

 

1. The visual representation of Table 3 is unappealing, and the removal of the first row may be beneficial.

 

Our response:

<> Thank you for the detailed review. We have deleted the first line..

 

2. The predicted image sizes of various models in Figure 11 are diminutive and uniform, warranting the highlighting of distinct segments of the models to enhance the readers' experience.

 

Our response:

<> Thank you for the detailed review. In view of the differences of the model, we have adopted a red border to mark it. The revised picture is as follows.

 

3. The bolding of the corresponding indicators in Table 7 should align with Table 6.

 

Our response:

<> Thank you for the detailed review. We have modified the indicators in the Table 7.

Loss Function

IoU

Acc

F-score

mIoU

mAcc

aAcc

MR

MP

MF

 

4. Furthermore, the article should furnish contextual information on the specific challenges encountered by the wetland ecosystem in Yancheng, Jiangsu.

 

Our response:

<> Thank you for the detailed review. We have added the background of coastal wetlands in Yancheng, Jiangsu.

Changes in precipitation and temperature can potentially affect the landscape pattern of coastal wetlands. The gradual rise in temperature, resulting in increased surface water evaporation, reduced precipitation, and inadequate water resource availability, has placed an additional burden on coastal wetlands in maintaining ecological equilibrium to a certain extent. Moreover, the growing population in Yancheng has contributed to the degradation of the coastal wetlands. The demand for ecological resources provided by wetlands has surged, placing immense pressure on these fragile ecosystems. Excessive development and utilization have further exacerbated the situation, leading to a significant decrease in the coastal wetland area.Consequently, considering the rapidly evolving wetland environment, the implementation of efficient and prompt detection methods becomes imperative.

 

 

Author Response File: Author Response.doc

Reviewer 4 Report

The manuscript entitled "Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Seformer" presents an application of the Segformer model in Sentinel-2 data to classify the coastal wetlands in China, a recently developed image segmentation model. The paper is well-written, with a clear methodology presentation, and provides a good application of the segformer model. The authors should see two minor points before considering the manuscript suitable for publication. Firstly, the evaluation metrics should be in the Method section. Finally, the authors should improve the discussions about the metrics obtained, especially for the river system (Is Acc the same as aAcc, in Table 7?).

Author Response

 

Manuscript ID: remotesensing-2451164

Paper Title: Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer

Authors: XuFeng Lin , YouWei Cheng , Gong Chen , WenJing Chen , Rong Chen , DeMin Gao , YinLong Zhang , YongBo Wu *

 

 

Dear Editors/Reviewers:

We would like to take this opportunity to thank you for your precious time and great efforts in handling the above referenced manuscript. Thanks also go to the anonymous reviewers for their valuable comments that helped to improve the quality of the paper. Overall the comments have been fair, encouraging and constructive. We have learned much from it. We have thought over the comments and revised the paper carefully.

 

Enclosed please find a summary of changes, a list of one-to-one responses to the comments of reviewers and the editor. Please kindly let us know if there are any additional comments/concerns. We are looking forward to hearing from you.

 

 

 

Sincerely,

Xufeng Lin, Zhongyuan Li, Wenjing Chen and Demin Gao  

july 12, 2023

 

 

 

Response to Reviewer ’s comments

 

1. Firstly, the evaluation metrics should be in the Method section. 

 

Our response:

<> Thank you for the detailed review. We have put the evaluation index in the method part.

 

2. Finally, the authors should improve the discussions about the metrics obtained, especially for the river system (Is Acc the same as aAcc, in Table 7?).

 

Our response:

<> Thank you for the detailed review. It may be that our statement has caused you some confusion. Here we use Acc to illustrate the segmentation accuracy of river system, and also to show the segmentation accuracy of different loss functions to the whole, so we use aAcc to show the segmentation accuracy of the whole, not the index for a single river..

 

 

Author Response File: Author Response.doc

Round 2

Reviewer 2 Report

The authors have not considered the comments. Please take the comments into consideration in detail. Some of the comments, like adding a discussion on UNet-based methods or improving the resolution of images are not incorporated in detail. The paper also has editorial issues such as tables going over the margin. Finally, I don't understand how come the results are the same each time the code is run. Any neural network such as a transformer is a non-convex function and each run should lead to a different result. Please apply the comments in detail for the next round.

The English can be followed but in terms of format, the paper needs further imprvements.

Author Response

 

Manuscript ID: remotesensing-2451164

Paper Title: Semantic segmentation of China's coastal wetlands based on Sentinel-2 and Segformer

Authors: XuFeng Lin , YouWei Cheng , Gong Chen , WenJing Chen , Rong Chen , DeMin Gao , YinLong Zhang , YongBo Wu *

 

 

Dear Editors/Reviewers:

We would like to take this opportunity to thank you for your precious time and great efforts in handling the above referenced manuscript. Thanks also go to the anonymous reviewers for their valuable comments that helped to improve the quality of the paper. Overall the comments have been fair, encouraging and constructive. We have learned much from it. We have thought over the comments and revised the paper carefully.

 

Enclosed please find a summary of changes, a list of one-to-one responses to the comments of reviewers and the editor. Please kindly let us know if there are any additional comments/concerns. We are looking forward to hearing from you.

 

 

 

Sincerely,

Xufeng Lin, Zhongyuan Li, Wenjing Chen and Demin Gao  

July 18  2023

 

 

 

Response to Reviewer ’s comments

 

1. The authors have not considered the comments. Please take the comments into consideration in detail. Some of the comments, like adding a discussion on UNet-based methods or improving the resolution of images are not incorporated in detail. The paper also has editorial issues such as tables going over the margin. Finally, I don't understand how come the results are the same each time the code is run. Any neural network such as a transformer is a non-convex function and each run should lead to a different result. Please apply the comments in detail for the next round. 

 

Our response:

<> Thank you for your valuable comments. We have examined the unet code carefully from beginning to end and found that there are indeed some parameter setting errors in the code that caused the unet model to fail to achieve the expected results. The code has been modified and rechecked. We have re-executed the experimental content and made changes in the predicted effect display. To address the editing issues, we have reformatted some tables and touched up the article. As you mentioned, the results of multiple runs of a general neural network code will vary. For the sake of reproducibility of the study, we fixed random seeds. After several experiments, the results are not guaranteed to be identical, but they are within a manageable margin of error (within 0.5%).

Author Response File: Author Response.doc