Coastal Zone Classification Based on U-Net and Remote Sensing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

thank you for your fascinating article.

Research LULC classification based on the U-Net deep learning model and high-resolution remotely sensed data is relevant for sustainable development under climate change. 

This study shows that ResNet50 is used as the feature extraction network of U-Net, and the coastal areas are accurately classified to evaluate the model's performance.

In common, additional data about estimation errors should improve this research. It is the main remark.

You can find comments and remarks below.

L. 2-4 – It is a long article title. Try shortening the title.

L. 39 –Authors should change the keywords “high resolution remotely sensed data” and “U-Net deep learning model” because the title and keywords should not have the same words. In the present version, the title contains the words “high resolution remotely sensed data” and “U-Net deep learning model”. Also, I recommend to use 6-9 keywords.

L. 47 – statistical

L. 92 – 2.1.1 Study …

L. 94 – Please, add “China” after “Guangdong Province”.

L. 103 – There are mistakes with the order numbers of references ([53,54]). It should be [47, 48].

L. 109 and Table 1 – Authors use definitions: forest land and woodland; water and water bodies; artificial ground and artificial surface; cultivated land and farmland. It is very important to know what do authors mean. For example, “forest” has many definitions in different countries. At the same time, readers can understand different objects"woodland" and "forest land". 

Figure 1 – Please, improve the figure. There are unclear coordinates. Also, it will be great to show the study area on a larger scale.

L. 122 – Could authors add a general workflow of the study?

Figure 2 – Please, improve the figure. There are unclear text, symbols and numbers.

Figure 3 – Please, improve the figure. There are unclear text, symbols and numbers.

L. 244 – Figure 5 doubles data from Table 2. Please, delete one of them.

L. 499-502 – Make sure reference #54 was correctly prepared.

Figure 4. Should be “Woodland”, “Farmland”

Figure 6. Should be “Woodland”, “Arable land” (or “Farmland”)

L. 244 – Figure 7 doubles data from Table 3. Please, delete one of them.

L.305-340 – Authors have to improve significantly the Discussion.

Comments on the Quality of English Language

Moderate editing of English language required

Author Response

Author’s Response to All the Comments

Dear Editor and Reviewers:

We are particularly grateful for your careful reading, and for giving us such constructive comments on this work! Those comments are all valuable and very helpful for revising and improving our manuscript, as well as the important guiding significance to our research. We have studied the comments carefully and have made revisions which we hope meet with approval.

According to the comments and suggestions, we have tried our best to improve the previous manuscript applsci-3111343 (“LULC Classification Based on U-Net Deep Learning Model and High Resolution Remotely Sensed Data”). Here is a summary of the major changes to the revised manuscript, and then answer the reviewer's questions one by one.

We have made significant adjustments and revisions to the content of the manuscript as requested. The main changes revisions include, (1) the Discussion part was revised to highlight the major contributions and the novelty of our study by comparing it to others; (2) the specific comments and suggestions were responded point by point; (3) the language and grammar of the article was polished; (4) The abstract part was significantly improved. All changes and author responses are marked in blue font in the manuscript.

Once again, we are particularly grateful for your careful reading and constructive comments. Thank you very much for your time. 

Best regards,

 

Pei Liu and all co-authors

 

2024-7-27

 

 

 

 

 

 

 

 

 

 

 

Author’s Response to the Comments of Reviewer #1

Dear authors, thank you for your fascinating article. Research LULC classification based on the U-Net deep learning model and high-resolution remotely sensed data is relevant for sustainable development under climate change. This study shows that ResNet50 is used as the feature extraction network of U-Net, and the coastal areas are accurately classified to evaluate the model's performance. In common, additional data about estimation errors should improve this research. It is the main remark. You can find comments and remarks below.

1. 2-4 – It is a long article title. Try shortening the title

Responses: Thank you very much for your suggestion. After careful consideration and careful deliberation, we decided to change the title to “Coastal Zone Classification Based on U-Net and Remote Sensing”. The revised title shortens the length of the text without changing the original meaning.

 

2. 39 –Authors should change the keywords “high resolution remotely sensed data” and “U-Net deep learning model” because the title and keywords should not have the same words. In the present version, the title contains the words “high resolution remotely sensed data” and “U-Net deep learning model”. Also, I recommend to use 6-9 keywords.

Responses: Thank you very much for your valuable comments and suggestions. The keywords of the manuscript were modified to “advanced remote sensing imaging; deep learning models; spectral and spatial information; feature extraction; coastal zone area; land use and cover classification; information interpretation technology”. The revised keywords are more accurate and specific, and cover the core content of the paper. They are more conducive to helping readers quickly understand the theme and content of the paper and facilitate the retrieval of related literature.

 

3. 47 – statistical

Responses: Thank you very much for pointing out this defect. This is a mistake caused by automatic line breaking in the office software. We have corrected this problem in the revised version.

 

4. 92 – 2.1.1 Study …

Responses: Thank you very much for your careful review and suggestions. We have revised the lowercase letters to uppercase letters. We apologize for the inconvenience caused by our carelessness.

 

5. 94 – Please, add “China” after “Guangdong Province”.

Responses: Thank you for your suggestions. We have added China after Guangdong Province.

 

6. 103 – There are mistakes with the order numbers of references ([53,54]). It should be [47, 48].

Responses: Thank you very much for your careful review. These two references [53,54] are newly added. And we have adjusted the order of references from 48 to 54. And during the process of manuscript revision, we added some new references, so the references here have been adjusted as [65,66].

 

7. 109 and Table 1 – Authors use definitions: forest land and woodland; water and water bodies; artificial ground and artificial surface; cultivated land and farmland. It is very important to know what do authors mean. For example, “forest” has many definitions in different countries. At the same time, readers can understand different objects "woodland" and "forest land".

Responses: Thank you for your comments. In the study of land use and land cover classification with remote sensing approach, different categories are defined based on their functions and physical characteristics. In our research, through field investigation and visualization interpretation, the research area is divided into 4 classes including water body, artificial surface, forest land, and farm land. Here the “forest land” means area dominated by trees includes both natural forests and plantations. “Water body” means area covered by water, such as rivers, lakes, wetlands, and artificial water bodies, such as reservoirs, ponds, etc. “Farm land” means land used for agricultural production, including arable land, orchards, vineyards, pastures, and agricultural facilities. Farm land is primarily used for growing crops, rising livestock, and aquaculture. “Artificial surface” means areas where the natural land cover has been replaced by human-made constructions, including urban and rural building, roads, airports, industrial areas, and other infrastructure. Artificial surfaces are primarily covered by materials such as concrete, asphalt, and bricks.

 

8. Figure 1 – Please, improve the figure. There are unclear coordinates. Also, it will be great to show the study area on a larger scale.

Responses: Thank you for your suggestions. We have updated figure 1 according to your suggestion. The updated figure 1 can be seen as follow.

Figure 1. Study area.

 

9. 122 – Could authors add a general workflow of the study?

Responses: Thank you for your suggestions. We created a flowchart presents the workflow for the coastal zone classification based on U-Net deep learning and GF-2 remotely sensed data framework employed in this study shown as figure 2.

 

Figure 2. Workflow of coastal classification framework

10. Figure 2 – Please, improve the figure. There are unclear text, symbols and numbers. Figure 3 – Please, improve the figure. There are unclear text, symbols and numbers.

Responses: Thank you for your suggestion. We have updated figure 2 and figure 3 as follow. And adjust the picture numbers to Figure 3 and Figure 4.

Figure 3. Structure of U-Net network.

Figure 4. Training accuracy, verification accuracy and loss value of U-Net model.

 

11. 244 – Figure 5 doubles data from Table 2. Please, delete one of them.

Responses: Thank you for your suggestion. Considering the need to express the experimental effects of different models more clearly and specifically, we choose to keep Table 2 and delete Figure 5.

 

12. 499-502 – Make sure reference #54 was correctly prepared.

Responses: Thank you very much for your careful review. Reference #54 are checked and updated.

 

13. Figure 4. Should be “Woodland”, “Farmland”, And Figure 6. Should be “Woodland”, “Arable land” (or “Farmland”)

Responses: Thank you very much for your valuable suggestions. We have carefully verified the description of LULC type, and have made the categories names in Figure 4 and Figure 6 consistent.

 

14. 244 – Figure 7 doubles data from Table 3. Please, delete one of them.

Responses: Thank you for your suggestion. Considering the need to express the experimental effects of different models more clearly and specifically, we choose to keep Table 3 and delete Figure 7.

 

15. 305-340 – Authors have to improve significantly the Discussion.

Responses: Thank you very much for your suggestions. We redesigned the structure of the results and discussion section and conducted a significantly analysis of the discussion section based on the research results and relevant references. The initial discussion section was structured as follow,

After a major revision, the construct of ‘Discussion’ section was updated as follow,

The detailed revision information is as follows,

4 Discussion

4.1 Advantages of U-Net deep learning models

In this research, a comprehensive comparison of U-Net model for coastal zone LUCC over the state-of-art deep learning models, such as SegNet, DeepLab v3+, as well as traditional machine learning models such as SVM and RF algorithms were performed. Experimental results demonstrated that U-Net has emerged as a powerful tool for land use and cover classification tasks due to its unique architecture. The design of U-Net includes a contracting path to capture context and a symmetric expanding path that enables precise localization, which is particularly beneficial for tasks requiring detailed boundary preservation [36]. While traditional machine learning such as SVM and RF, requires extensive feature engineering to perform well, whereas deep learning models like U-Net can automatically learn relevant features from raw data. This makes SVM less flexible and more dependent on the quality of the engineered features [55, 56]. Our experiment results shown that because of the ability of automatic process of feature extraction and multiple layer transition, with a OA better than 85%, the selected three deep learning algorithms outperform the two traditional statistical learning models, this is consistent with Li's research conclusions in the Shenzhen area. [57]. Although Deeplab's calculation accuracy is closest to the U-Net model, however, the results of relevant literature show that DeepLab v3+ is more complex and computationally intensive than U-Net [58]. U-Net stands out for its high accuracy and detailed segmentation, particularly when dealing with complex boundaries and limited training data. SegNet offers computational efficiency but may sacrifice some detail. DeepLab v3+ provides state-of-the-art performance but at the cost of increased complexity and computational requirements. SVM and Random Forest are robust traditional methods but require extensive feature engineering and may not perform as well in high-dimensional data scenarios without deep learning’s automatic feature learning capabilities.

4.2 Benefit of spectral and spatial features

 U-Net with spectral and spatial features combination method achieves the best classification performance (OA over 93%) by using the GF-2 images of Shuangyue Bay, Guangdong province as the training set, where the surface cover is complex. The analysis results show that NDVI is more sensitive to vegetation characteristics. This is mainly because NDVI uses green plants to absorb strongly in the red-light band, and its high reflectance in the near-infrared increases the spectral response difference between vegetation and other ground objects. In addition, texture features have improved the classification accuracy of all features. This is mainly because texture feature, as a regional feature, makes full use of image information to describe the spatial distribution of each pixel in the image. Compared with other features, it can better consider both macroscopic properties and fine structure. The contrast feature is more sensitive to water bodies and farm land. This is because the contrast feature combines the gray levels of low-frequency pixels in the original image and stretches the gray levels of high-frequency pixels to highlight the details of the image and increase the gap between water bodies, cultivated land, and other features difference. Especially, the fusion of texture, NDVI, and contrast features achieves the best classification effect. This is mainly because each feature has a different contribution to different categories. When these features are combined, their advantages can be fully utilized.

4.3 Key bottlenecks and future directions

Deep learning methods require a large amount of data to train the model to achieve excellent performance, but the acquisition and labeling of data set is costly and difficult. In addition, it is found that although spectral and spatial features added, the increase in classification accuracy of artificial surface is very limited. This indicates that the selected features are not very sensitive to artificial surface. Therefore, constructing building index features is expected to solve this problem. However, since most high-resolution remote sensing images only have red, green, blue and near-infrared channels, it is difficult to build a normalized building index that requires a mid-infrared channel [59]. Therefore, using hyperspectral remote sensing data or combining high-resolution satellite images with medium-resolution remote sensing images to make up for the shortcomings of optical resolution satellite images in extracting building land is an effective way to solve this problem.

References:

[36] Ronneberger, O.; Fischer, P.; Brox, T. U-net：Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Comput‐er-Assisted Intervention, Munich, Germany, 5-9 October 2015;  pp. 234-241.

[55] Pal, M. and  Mather, P. M. Support Vector Machines for Classification in Remote Sensing. International Journal of Remote Sensing, 2005, 26(5), 1007-1011.

[56] Belgiu, M. and Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS Journal of Photogrammetry and Remote Sensing, 2016, 114, 24-31

[57] Z. Li; B. Chen; S. Wu, et al. Deep learning for urban land use category classification: A review and experimental assessment, Remote Sensing of Environment, 2024, 311,114290

[58] Chen, LC., Zhu, Y., Papandreou, G., Schroff, F., Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. Computer Vision, 2018,arXiv:1802.02611. https://doi.org/10.1007/978-3-030-01234-2_49

[59] Zhang, Q.and Seto, K. C. Mapping urbanization dynamics at regional and global scales using multitemporal DMSP/OLS nighttime light data. Remote Sensing of Environment, 2011, 115(9), 2320-2329

 

Once again, we are particularly grateful for your careful reading and constructive comments. Thanks very much for your time.

Best regards,

Pei Liu and all co-authors

2024-7-27

 

 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents an analysis of the land use/land cover classification of a coastal zone using the U-Net method. The topic is worthy of research; however, it has several details that need to be addressed. The authors should address several minor changes before it is considered for publication.

General Comments

C1. The main concern is that the novelty of the research is not fully clear. If such novelty is not clearly highlighted, the risk is that the manuscript looks more a simple case study rather than a research paper.

C2. What are the major contributions of this study? should be carefully mentioned in the discussion section.

C3. There is an absence of a discussion where detailed comparisons of the results with those of other researchers in the field are made. I recommend including some relevant references to enhance the discussion of the novelty of your study by comparing it to others. This will provide a stronger framework for understanding the implications of your findings and highlight the unique contributions of your research.

*The answer to these question should be reflected in the manuscript.*

Specific comments

Line 32: What are “these obvious geographical advantages”? Mention and detail.

Line 32: Change “sup-port” to “support”.

Line 37: The statement should be supported by relevant literature.

Line 38: Change “spatial resolution…” to “spatial and spectral resolution…”

Line 41: What is that level of precision, in which studies?

Line 57: Detail the “obvious advantages”.

Line 70-71: What are those “obvious shortcomings”. Mention and detail.

Line 93-94: Provide more details of the selection of the “typical coastal zone” study area.

Line 96-97: Please add website, add the accessed-on time as day month year.

Line 97: Why was January 26, 2017 selected, and not some more recent date (e.g. 2023, 2024).

Line 98-103: What is the radiometric resolution?

Line 107-108: To ensure the veracity of the results, it is important that the authors mention how the radiometric calibration and atmospheric correction process was carried out, among other relevant details. Detail the parameters of these processes.

Line 112: Delete “OOM”.

Line 119: Which version?

Line 141: The figure has very low quality and the details are not visible. Improve extensively. Perform this throughout the manuscript

Line 160-162: Include a more detailed explanation of how texture features are calculated using GLCM.

Line 191-193: Why were OA, kappa and F1-Score selected as accuracy assessment indices?

Line 218: The figure has very low quality and details are not visible. Improve.

Line 341: Conclusions should reflect the limitations of the study and propose future areas of research based on the main findings and limitations.

Author Response

Author’s Response to All the Comments

Dear Editor and Reviewers:

We are particularly grateful for your careful reading, and for giving us such constructive comments on this work! Those comments are all valuable and very helpful for revising and improving our manuscript, as well as the important guiding significance to our research. We have studied the comments carefully and have made revisions which we hope meet with approval.

According to the comments and suggestions, we have tried our best to improve the previous manuscript applsci-3111343 (“LULC Classification Based on U-Net Deep Learning Model and High Resolution Remotely Sensed Data”). Here is a summary of the major changes to the revised manuscript, and then answer the reviewer's questions one by one.

We have made significant adjustments and revisions to the content of the manuscript as requested. The main changes revisions include, (1) the Discussion part was revised to highlight the major contributions and the novelty of our study by comparing it to others; (2) the specific comments and suggestions were responded point by point; (3) the language and grammar of the article was polished; (4) The abstract part was significantly improved. All changes and author responses are marked in blue font in the manuscript.

Once again, we are particularly grateful for your careful reading and constructive comments. Thank you very much for your time. 

Best regards,

 

Pei Liu and all co-authors

 

2024-7-27

 

 

 

 

 

 

 

 

 

 

 

Author’s Response to the Comments of Reviewer #2

The paper presents an analysis of the land use/land cover classification of a coastal zone using the U-Net method. The topic is worthy of research; however, it has several details that need to be addressed. The authors should address several minor changes before it is considered for publication.

1. The main concern is that the novelty of the research is not fully clear. If such novelty is not clearly highlighted, the risk is that the manuscript looks more a simple case study rather than a research paper.

Responses: Thank you for your valuable comments and suggestions. In this study, we selected Gaofen-2, China's high-resolution satellite remote sensing images, as the main data source to test and verify the advantages of U-Net deep learning in land use coverage classification in coastal zone. Classification outcomes were compared and verified with the state-of-art deep learning algorithms such as SegNet, Deeplab v3+ and traditional machine learning algorithms SVM and RF algorithms. The innovation of this research can be highlighted as, (1) scenes of Gaofen-2 (GF-2) high resolution remotely sensed data over Shuangyue Bay, a typical coastal zone in Guangdong Province were tested; (2) coastal land use and land cover classification results were compared with several most popular deep learning models such as SegNet, DeepLab v3+ as well as some advanced statistical learning models such as Support Vector Machine (SVM), Random Forest (RF); (3) the contribution of artificial extracted texture, NDVI, contrast features for U-Net deep learning model was examined. 

 

2. What are the major contributions of this study? should be carefully mentioned in the discussion section. There is an absence of a discussion where detailed comparisons of the results with those of other researchers in the field are made. I recommend including some relevant references to enhance the discussion of the novelty of your study by comparing it to others. This will provide a stronger framework for understanding the implications of your findings and highlight the unique contributions of your research.

Responses: Thank you for your valuable comments and suggestions. We have reorganized and carefully revised the Discussion section to better express the academic contribution of this study. The major contribution of this study can be summarized as (1) the advantage of deep learning approach especially the U-Net deep learning model for LULC classification over coastal zone based on high resolution remote sensing images was well tested; (2) The spectral and spatial combination method and the contributions of spectral and spatial features of GF-2 data for different land cover types were analyzed. We also including some relevant references to enhance the discussion of the novelty of your study by comparing it to others. More revision details can be described as follows,

4.1 Advantages of U-Net deep learning models

In this research, a comprehensive comparison of U-Net model for coastal zone LUCC over the state-of-art deep learning models, such as SegNet, DeepLab v3+, as well as traditional machine learning models such as SVM and RF algorithms were performed. Experimental results demonstrated that U-Net has emerged as a powerful tool for land use and cover classification tasks due to its unique architecture. The design of U-Net includes a contracting path to capture context and a symmetric expanding path that enables precise localization, which is particularly beneficial for tasks requiring detailed boundary preservation [52]. While traditional machine learning such as SVM and RF, requires extensive feature engineering to perform well, whereas deep learning models like U-Net can automatically learn relevant features from raw data. This makes SVM less flexible and more dependent on the quality of the engineered features [75, 76]. Our experiment results shown that because of the ability of automatic process of feature extraction and multiple layer transition, with an OA better than 85%, the selected three deep learning algorithms outperform the two traditional statistical learning models, this is consistent with Li's research conclusions in the Shenzhen area [77]. Although Deeplab's calculation accuracy is closest to the U-Net model, however, the results of relevant literature show that DeepLab v3+ is more complex and computationally intensive than U-Net [78]. U-Net stands out for its high accuracy and detailed segmentation, particularly when dealing with complex boundaries and limited training data. SegNet offers computational efficiency but may sacrifice some detail. DeepLab v3+ provides state-of-the-art performance but at the cost of increased complexity and computational requirements. SVM and Random Forest are robust traditional methods but require extensive feature engineering and may not perform as well in high-dimensional data scenarios without deep learning’s automatic feature learning capabilities.

4.2 Benefit of spectral and spatial features

 U-Net with spectral and spatial features combination method achieves the best classification performance (OA over 93%) by using the GF-2 images of Shuangyue Bay, Guangdong province as the training set, where the surface cover is complex. The analysis results show that NDVI is more sensitive to vegetation characteristics. This is mainly because NDVI uses green plants to absorb strongly in the red-light band, and its high reflectance in the near-infrared increases the spectral response difference between vegetation and other ground objects. In addition, texture features have improved the classification accuracy of all features. This is mainly because texture feature, as a regional feature, makes full use of image information to describe the spatial distribution of each pixel in the image. Compared with other features, it can better take into account both macroscopic properties and fine structure. The contrast feature is more sensitive to water bodies and farm land. This is because the contrast feature combines the gray levels of low-frequency pixels in the original image and stretches the gray levels of high-frequency pixels to highlight the details of the image and increase the gap between water bodies, cultivated land, and other features difference. Especially, the fusion of texture, NDVI, and contrast features achieves the best classification effect. This is mainly because each feature has a different contribution to different categories. When these features are combined, their advantages can be fully utilized.

4.3 Key bottlenecks and future directions

Deep learning methods require a large amount of data to train the model to achieve excellent performance, but the acquisition and labeling of data set is costly and difficult. In addition, it is found that although spectral and spatial features added, the increase in classification accuracy of artificial surface is very limited. This indicates that the selected features are not very sensitive to artificial surface. Therefore, constructing building index features is expected to solve this problem. However, since most high-resolution remote sensing images only have red, green, blue and near-infrared channels, it is difficult to build a normalized building index that requires a mid-infrared channel [79]. Therefore, using hyperspectral remote sensing data or combining high-resolution satellite images with medium-resolution remote sensing images to make up for the shortcomings of optical resolution satellite images in extracting building land is an effective way to solve this problem.

[52] Ronneberger, O.; Fischer, P.; Brox, T. U-net：Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Comput‐er-Assisted Intervention, Munich, Germany, 5-9 October 2015;  pp. 234-241.

[75] Pal, M. and  Mather, P. M. Support Vector Machines for Classification in Remote Sensing. International Journal of Remote Sensing, 2005, 26(5), 1007-1011.

[76] Belgiu, M. and Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS Journal of Photogrammetry and Remote Sensing, 2016, 114, 24-31

[77] Z. Li; B. Chen; S. Wu, et al. Deep learning for urban land use category classification: A review and experimental assessment, Remote Sensing of Environment, 2024, 311,114290

[78] Chen, LC., Zhu, Y., Papandreou, G., Schroff, F., Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. Computer Vision, 2018, arXiv:1802.02611. https://doi.org/10.1007/978-3-030-01234-2_49

[79] Zhang, Q.and Seto, K. C. Mapping urbanization dynamics at regional and global scales using multitemporal DMSP/OLS nighttime light data. Remote Sensing of Environment, 2011, 115(9), 2320-2329

 

3. Line 32: What are “these obvious geographical advantages”? Mention and detail.

Responses: Thank you for your comments. Due to its unique location at the junction of land and sea, the coastal zone has obvious geographical advantages, covering multiple levels such as ecology, economy and society. Coastal zones often host diverse ecosystems such as estuaries, mangroves, coral reefs, and wetlands, which provide critical habitats for various species [1]. Coastal zones support major industries such as fishing, tourism, and shipping [2]. Coastal areas are often sites of significant cultural and historical importance, attracting tourism and promoting cultural heritage. The aesthetic and recreational value of coastal zones contributes to the quality of life for residents and visitors.

References:  

[1] Costanza, R., d'Arge, R., de Groot, et al. The value of the world's ecosystem services and natural capital. Nature, 387(6630), 1997, 253-260.

[2] Barbier, E. B. Progress and challenges in valuing coastal and marine ecosystem services. Review of Environmental Economics and Policy, 6(1), 2012, 1-19

 

4. Line 32: Change “sup-port” to “support”.

Responses: Thank you very much for your careful review. The word “sup-port” was updated as “support”.

 

5. Line 37: The statement should be supported by relevant literature.

Responses: Thank you for your suggestion. We have added the following highly relevant references.

[9] Micallef, A. and Williams, A. T. Theoretical strategy considerations for beach management. Ocean & Coastal Management, 2002, 45(4-5), 261-275.

[10] Kuleli, T., Guneroglu, A., Karsli, F., et al. Automatic detection of shoreline change on coastal Ramsar wetlands of Turkey. Ocean Engineering, 2011, 38(10), 1141-1149.

[11] Seto, K. C. and Fragkias, M. Quantifying spatiotemporal patterns of urban land-use change in four cities of China with time series landscape metrics. Landscape Ecology, 2005, 20, 871-888.

 

6. Line 38: Change “spatial resolution…” to “spatial and spectral resolution…”

Responses: Thank you very much for your suggestion. The “spatial resolution” has been changed to “spatial and spectral resolution”.

 

7. Line 41: What is that level of precision, in which studies?

Responses: Thank you for your question. The accuracy of traditional human interpretation of remote sensing image classification usually depends on many factors, such as the experience of the interpreter, image resolution, clarity of ground features, and complexity of the classification system. In general, the accuracy of visual interpretation is usually between 70% and 95%. This paper discussed the use of medium spatial resolution satellite data for land cover classification over large areas, noting that visual interpretation can achieve 80% to 90% accuracy in some cases [1]. While Harris and Ventura discussed research on combining geographic data with remote sensing imagery to improve classification accuracy in urban areas, highlighting that visual interpretation can achieve an accuracy of about 80% in medium-resolution imagery [2].

• Franklin, S. E., and Wulder, M. A. Remote sensing methods in medium spatial resolution satellite data land cover classification of large areas. Progress in Physical Geography, 2002, 26(2), 173-205.
• Harris, P. M., and Ventura, S. J. The integration of geographic data with remotely sensed imagery to improve classification in an urban area. Photogrammetric Engineering and Remote Sensing, 1995, 61(8), 993-998.

 

8. Line 57: Detail the “obvious advantages”.

Responses: Thank you for your comments. Deep learning techniques offer several advantages in remote sensing image classification compared to traditional methods. These advantages include improved accuracy, the ability to handle large and complex datasets, automatic feature extraction, and the capacity to learn from multi-source and multi-temporal data. For instance, Zhu et al [1] discussed the state-of-the-art in deep learning for remote sensing and highlights the significant improvements in classification accuracy brought about by deep learning methods. Ma et al [2] reviewed deep learning applications in remote sensing and emphasizes the ability of deep learning models to handle large and complex datasets. This study illustrates how convolutional neural networks (CNNs) can automatically extract relevant features from hyperspectral images, eliminating the need for manual feature engineering [3].

• Zhu, X. X., Tuia, D., Mou, L., et al. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE Geoscience and Remote Sensing Magazine, 2017, 5(4), 8-36
• Ma, L., Liu, Y., Zhang, X., et al. Deep learning in remote sensing applications: A meta-analysis and review. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 152, 166-177
• Chen, Y., Jiang, H., Li, C., et al. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(10), 6232-6251

 

9. Line 70-71: What are those “obvious shortcomings”. Mention and detail.

Responses: Thank you for your comments. While Fully Convolutional Networks (FCNs) have shown great promise in remote sensing image classification, they also have several shortcomings. For the tasks such as remotely sensed data classification, FCNs require substantial computational resources, especially for high-resolution remote sensing images, which can be a limitation when processing large datasets or deploying models on resource-constrained devices[1]. This algorithm always struggle to capture long-range dependencies and contextual information in remote sensing images, which can be crucial for accurate classification of complex scenes [2]. The shortcomings also include it often face challenges in effectively dealing with objects of varying scales, which is common in remote sensing imagery [3]. During the training stage, FCNs can overfit to the training data, especially when the training dataset is small or lacks diversity. This can result in poor generalization to new, unseen data [4]. So FCNs typically require large labeled datasets for training, which can be difficult and expensive to obtain in the context of remote sensing [5]. Finally, FCNs can be sensitive to noise and variability in remote sensing images, which can affect classification accuracy [6].

• Zhang, C., Wei, S. and Zhang, Y. A Review on Image Segmentation Techniques with Remote Sensing Perspective. IEEE Geoscience and Remote Sensing Magazine, 2018, 6(3), 61-77.
• Fu, G., Liu, C., Zhou, R., et al. Classification for High Resolution Remote Sensing Imagery Using a Fully Convolutional Network. Remote Sensing, 2017, 9(5), 498.
• Masi, G., Cozzolino, D., Verdoliva, L., and Scarpa, G. Pansharpening by Convolutional Neural Networks. Remote Sensing, 2016, 8(7), 594.
• Kemker, R., Luu, R., & Kanan, C. Low-Shot Learning for the Semantic Segmentation of Remote Sensing Imagery. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(1), 329-340.
• Volpi, M., & Tuia, D. Deep multi-task learning for a geographically-regularized semantic segmentation of aerial images. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 144, 48-60
• Makantasis, K., Karantzalos, K., Doulamis, A., & Doulamis, N. Deep Supervised Learning for Hyperspectral Data Classification Through Convolutional Neural Networks. IEEE Geoscience and Remote Sensing Letters, 2015, 13(1), 5-9.

 

10. Line 93-94: Provide more details of the selection of the “typical coastal zone” study area.

Responses: Thank you for your comments. Shuangyue Bay is known for its distinctive dual crescent-shaped bays, which create a unique and picturesque coastal landscape. This characteristic makes it an exemplary case for studying coastal geomorphology [1]. Like many coastal zones, Shuangyue Bay faces environmental pressures from human activities such as tourism, fishing, and urbanization. These factors make it a representative area for studying the impact of human activities on coastal environments and for developing management strategies [2].

[1] Li, M., Wang, H., and Zhang, W. The Geographical Characteristics and Tourism Resources of Shuangyue Bay, Guangdong. Geographical Research, 2014, 33(4), 789-797.

[2] Liu, Z., Zhang, J., and Huang, W. Preliminary Study on the Impact of Human Activities on the Coastal Environment of Shuangyue Bay, Guangdong Province. Marine Sciences, 2015, 40(6), 112-120.

We added this supplementary content to the subsection of study area.

 

11. Line 96-97: Please add website, add the accessed-on time as day month year.

Responses: Thank you for your suggestions. You can query and obtain GF-2 satellite remote sensing images from the following website (https://data.cresda.cn/#/2dMap). Since the high-resolution satellite remote sensing images are not currently available to the public for download, a confidentiality agreement is required in most cases. The data used in this study were obtained through a proprietary request. We have added the data query and download URLs, as well as the acquisition date, to the main body of the paper and marked them in blue font.

 

12. Line 97: Why was January 26, 2017 selected, and not some more recent date (e.g. 2023, 2024).

Responses: Thank you for your comments. This is because the distribution of high-resolution satellite images, such as GF-2, is based on provinces in China. And the dataset used in this study was due to a special project needs at that time. On January 26, 2017, we conducted field data collection in the study area and applied for GF-2 satellite images of the same period. In recent years, such as 2023 and 2024, we did not perform field work any more in the Guangdong province, so we could not apply for GF-2 data of that region.

 

13. Line 98-103: What is the radiometric resolution?

Responses: Thank you for your question. The radiometric resolution of GF-2 satellite image is 10 bits. We have also added this information to the manuscript.

 

14. Line 107-108: To ensure the veracity of the results, it is important that the authors mention how the radiometric calibration and atmospheric correction process was carried out, among other relevant details. Detail the parameters of these processes.

Responses: Thank you for your suggestions. We use the Apply Gain and Offset tool provided by ENVI to perform radiometric calibration on multispectral data. During this preprocess we input the entire scene image, shown as figure 1.

Figure 1. GF2 remote sensing image before atmospheric correction

The absolute calibration coefficient of GF2 remote sensing image in 2017 is that PAN channel (gain = 0.1503, bias = 0), B1 channel (gain = 0.1193, bias=0), B2 channel (gain = 0.1530, bias = 0), B3 channel (gain = 0.1424, bias = 0), B4 channel (gain = 0.1569, bias = 0). The resulting image after calibration of GF-2 is shown in the figure 2.

Figure 2. Calibration results of selected GF-2 image

After radiometric calibration, the FLASSH module is used to perform atmospheric correction on remote sensing data. The relevant parameter settings for atmospheric correction of the selected data are shown in Figure 3 (a) – Figure 3 (c).

(a) Atmospheric correction parameters

(b) Multispectral Setting

(c) Advanced parameters Setting

Figure 3. Parameter settings of atmospheric correction

The result after atmospheric correction is shown in Figure 4.

Figure 4. GF2 remote sensing image after atmospheric correction

 

15. Line 112: Delete “OOM”.

Responses: Thank you for your comments. The abbreviation OOM has been deleted.

 

16. Line 119: Which version?

Responses: Thank you for your question. During the experiment, we have the university version of MATLAB® R2020a.

 

17. Line 141: The figure has very low quality and the details are not visible. Improve extensively. Perform this throughout the manuscript

Responses: Thank you for your suggestions. Initially, we compressed the images to control the file size and facilitate typesetting. Now we have replaced all the images throughout the manuscript with the original images, hoping that the clarity and quality of the images can meet the requirements.

 

18. Line 160-162: Include a more detailed explanation of how texture features are calculated using GLCM.

Responses: Thank you for your suggestions. The texture features calculated using GLCM including preprocess the GF-2 remotely sensed data, convert the image to grayscale, define the angle parameter. In our experiment, the preprocess includes radiometric correction, geometric correction and resampling to ensure the images; GLCM was applied to each band of the four channels of multispectral GF-2 data to calculate textural features separately.

In this research, we extracted five texture features include contrast (CON), Correlation (COR), Angular Second Moment (ASM), Mean, Entropy (ENT) using GF-2 RGB channels with four angle directions (0°, 45°, 90°, 135°).

The  measures the local variations in the GLCM matrix, so which reflects image clarity and texture depth. The calculation formula of CON can be described as equation (1),

(1)

where,  and  means the row and column indices in the GLCM,  means the normalized frequency of the co-occurrence of gray levels i and j in the GLCM.

The  measures how correlated a pixel is to its neighbor over the whole image, which reflects the correlation of local grayscale of the image, The calculation formula of  can be described as equation (2),

 

(2)

Where, ,  and  means the row and column indices in the GLCM,  means the normalized frequency of the co-occurrence of gray levels i and j in the GLCM.  is the mean gray level of row  in the GLCM,  is the mean gray level of column j in the GLCM,  and  is the standard deviation of the gray levels in row  and column  respectively.

ASM measures the uniformity or texture uniformity of the image, which also known as Energy and reflects the uniformity of the image grayscale distribution and the thickness of texture. The calculation formula of ASM can be described as equation (3),

 

(3)

where,  and  means the row and column indices in the GLCM,  means the normalized frequency of the co-occurrence of gray levels i and j in the GLCM.

Mean is the average value of the intensities in the GLCM, which reflects the brightness of the image tone. The calculation formula of Mean can be described as equation (4),

 

(4)

where,  and  means the row and column indices in the GLCM,  means the normalized frequency of the co-occurrence of gray levels i and j in the GLCM.

ENT measures the randomness or complexity of the texture in the image and which reflects the randomness measurement of the amount of information contained in an image. The calculation formula of ENT can be described as equation (5),

 

(5)

where,  and  means the row and column indices in the GLCM,  means the normalized frequency of the co-occurrence of gray levels i and j in the GLCM. Log is the logarithm function, typically base 2 or natural logarithm.

 

19. Line 191-193: Why were OA, kappa and F1-Score selected as accuracy assessment indices?

Responses: Thank you for your question. In remote sensing classification tasks, it is crucial to evaluate the performance of classification algorithms accurately. Overall Accuracy (OA), Kappa coefficient, and F1-Score are commonly used metrics for this purpose [1,2]. Some reasons for the selection of OA can be summarized as (1) OA is easy to understand and calculate; (2) OA can give a quick overview of the classifier’s performance; (3) OA is a standard metric in many fields, making results easily comparable across studies. Some reasons for the selection of Kappa are including (1) kappa considers the agreement concurring by chance; (2) kappa is less sensitive to class imbalance; (3) values of kappa closer to 1 indicating better agreement than expected by chance, and values of kappa below than 0 indicating worse than random performance. The main reason for choosing the F1 parameter is because that (1) F1 score provides a single metric that balances both false positives and false negatives; (2) F1 score is useful in cases of class imbalance by considering both the precision and recall; (3) F1 score offers detailed performance insights.

[1] Stehman, S. V. Selecting and interpreting measures of thematic classification accuracy. Remote Sensing of Environment, 1997, 62(1), 77-89.

[2] Powers, D. M. Evaluation: From precision, recall and F-measure to ROC, informed, markedness and correlation. Journal of Machine Learning Technologies, 2011, 2(1), 37-63.

 

20. Line 218: The figure has very low quality and details are not visible. Improve.

Responses: Thank you for your suggestions. Initially, we compressed the images to control the file size and facilitate typesetting. Now we have replaced all the images throughout the manuscript with the original images, hoping that the clarity and quality of the images can meet the requirements.

 

21. Line 341: Conclusions should reflect the limitations of the study and propose future areas of research based on the main findings and limitations.

Responses: Thank you for your valuable comments and suggestions. After careful consideration, we have restructured and rewritten the conclusion section as follow,

Based on GF-2 remote sensing images, this paper presents an analysis of the land use/land cover classification of a coastal zone using the U-Net method. Through research, the effectiveness of GLCM texture features, NDVI, and contrast features in the U-Net model was verified, and a strategy based on high-resolution remote sensing images was constructed to obtain a LULC classification solution suitable for coastal areas. The comprehensive analysis of the experimental results demonstrated that (1) The U-Net classification algorithm that combines texture, NDVI and contrast features proposed in this paper can greatly improve the classification accuracy. (2) The model proposed in this paper can effectively classify artificial surface, forest land, farm land, and water bodies with a high performance in the classification of coastal features (classification accuracy reaches 93.65%), and it can better meet the requirements of practical applications.

While, there are still some shortcomings that need to be improved. The limitations of coastal zone classification based on GF-2 remote sensing data and U-Net model are mainly reflected in the following points, (1) GF-2 has limited spectral bands, which might not capture all necessary information for classification between various coastal features, especially those with similar spectral signatures; (2) U-Net model requires a substantial amount of high quality annotated training data, which can be labor-intensive and costly to produce; (3) Deep learning models trained on one coastal region might not perform well in another region due to differences in environmental conditions; (4) Coastal zones may have imbalanced classes, with some land cover types being underrepresented. This can lead to biases in the model, where less frequent classes are not classified accurately.

To overcome the above shortcomings and disadvantage, it is suggested that future research can be improved in the following aspects, (1) Integrated GF-2 data with other high-resolution and hyperspectral datasets to enhance spectral and temporal resolution; (2) Explore hybrid architectures that combine U-Net with other models (e.g., LSTM for temporal data) to improve performance in dynamic environments; (3) Use synthetic data generation techniques (e.g., GANs) to create additional training samples, especially for underrepresented classes; (4) Investigate semi-supervised learning approaches to leverage large amounts of unlabeled data, reducing the dependency on annotated data.

 

Once again, we are particularly grateful for your careful reading and constructive comments. Thanks very much for your time.

Best regards,

Pei Liu and all co-authors

2024-7-27

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

thank you for improving the article.

It would be great if you adapted the size of the Figures for better reading. 

Comments on the Quality of English Language

Moderate editing of the English language required

Author Response

Author’s Response to All the Comments

Dear Editor and Reviewers:

We are particularly grateful for your careful reading. And thank you for your further suggestions on revising the paper. We have studied the comments carefully and have made revisions which we hope meet with approval.

According to the comments and suggestions, we have tried our best to improve the previous manuscript applsci-3111343 (“Coastal Zone Classification Based on U-Net and Remote Sensing”). We have made revisions to the content of the manuscript as requested.

Once again, we are particularly grateful for your careful reading and constructive comments. Thank you very much for your time.  

Best regards,

 

Pei Liu and all co-authors

 

2024-8-1

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Author’s Response to the Comments of Reviewer #1

Comments and Suggestions for Authors

Dear authors,

Thank you for improving the article.

It would be great if you adapted the size of the Figures for better reading.

Responses: Thank you very much for your suggestion. We adjusted the size of the Figures in the paper one by one to make it more comfortable to read and provide a better reading experience.

 

Comments on the Quality of English Language

Moderate editing of the English language required

Responses: Thank you very much for your valuable comments. We have asked a professional language editing agency to polish the entire paper and attach an editorial certification. We also submitted a language editing comparison report in the Supplementary File(s),  hoping to get your approval.

Once again, we are particularly grateful for your careful reading and constructive comments. Thanks very much for your time.

Best regards,

Pei Liu and all co-authors

2024-8-1

 

 

Author Response File: Author Response.docx