The Effect of Negative Samples on the Accuracy of Water Body Extraction Using Deep Learning Networks

Round 1

Reviewer 1 Report

The paper "The Effect of Negative Samples on the Accuracy of Water body Extraction using Deep Learning networks" presents the usage of Deep Learning applied onto water bodies and cloud shadows in order to determine water body extraction accuracy. The results seem promising according to the obtained overall accuracy, mean intersection over union and Kappa parameters.

 

- Originality/Novelty: The paper is novel as it uses deep learning to determine water body extraction.

 

- Significance: The results of the research are interpreted properly, the fine tuning of model parameters being performed.

 

- Quality of Presentation: The article is written appropriately, respecting the logical succession of sections. Data and analyses are presented graphically and inside tables. The results were outlined using high standards, the advantages of the machine learning method being very clear.

 

The abbreviations OA and mIoU should be explained in the abstract, because they appear for the first time.

 

Please avoid the usage of pronouns like "we" or "I" in a journal paper.

 

The conclusions section has to include further details about future work.

 

How can an even better accuracy be obtained?

 

Arrange the text and figure 4 such that it will fit on page 8. For the current version, page 8 has just a paragraph and the rest is totally empty. Same observation for pages 12, 15.

 

Do not leave the discussion chapter title at the bottom on page 17 when its content is on the next page.

 

How does light influence the Sentinel-2 image? Which are the accuracy value changes for morning, noon, afternoon, and evening conditions?

 

- Scientific Soundness: The paper offers enough details to allow the reproduction of the results, but it needs to have a higher scientific value.

 

- Interest to the Readers: The content of the article would surely interest the readers of the Remote Sensing journal, and not only them. 

 

- Overall Merit:  The findings and their implications should be discussed in the broadest context possible and limitations of the work should also be further highlighted. Please present the novelty as compared to the previously published papers.

 

- English Level: The level of English language is advanced. Through the entire paper, the language was appropriate and understandable, being easy to follow the flow from the beginning.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Overview

The paper at hand proposes a deep learning approach for water body extraction from remote sensing images. The algorithm is evaluated regarding its capability of discerning water bodies from cloud shadows. For this purpose, a new dataset is created and exploited to evaluate the proposed deep neural network's performance.

Comments

1) Please rephrase: (lines 42-43) "Semantic segmentation networks in deep learning are mainly used networks in water body extraction." The authors probably imply that Deep autoencoder networks exploited in semantic segmentation tasks are mainly used also for water body extraction.

2) Line 43: Please refer also to SegNet in semantic segmentation networks

• Badrinarayanan, Vijay, Alex Kendall, and Roberto Cipolla. "Segnet: A deep convolutional encoder-decoder architecture for image segmentation." IEEE transactions on pattern analysis and machine intelligence 39.12 (2017): 2481-2495.
• Balaska, Vasiliki, et al. "Enhancing satellite semantic maps with ground-level imagery." Robotics and Autonomous Systems 139 (2021): 103760.

3) Line 57: "non-cloud images"

4) Line 59: "when the networks are trained with..."

5) Line 66: "Considering that deep learning..."

6) Line 66-71: the proposed work investigates the very interesting topic of dataset split and negative samples impact in deep learning-based classification performance of water bodies extraction. The above fact is mainly realized as the challenge of avoiding over-fitting in the deep feature space that is the input of the classification layer in a deep neural network. Please, refer to the initial study referring to this specific challenge in deep learning to further demonstrate the impact of the proposed study.

• Kansizoglou, Ioannis, Loukas Bampis, and Antonios Gasteratos. "Deep feature space: A geometrical perspective." IEEE Transactions on Pattern Analysis and Machine Intelligence 44.10 (2021): 6823-6838.

7) Line 71-73: please rephrase the entire sentence. Currently, it is very difficult to understand its meaning.

Possible suggestion: “We thus designed two groups of training datasets, including a) a training set that only contains labeled water bodies images and b) a second one containing labeled images of both water bodies and cloud shadows.”

8) Line 115-116: I believe the authors should refer to the initial and well-established implementation of Transformer based on self-attention:

• Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30, 2017.

9) Line 111-112: “The Swin Transformer”, Ref. [34] can be placed here.

10) Line 126-127: a reference of the Shifted-Window Multihead Self-Attention method should be placed

• Vaswani, Ashish, et al. "Scaling local self-attention for parameter efficient visual backbones." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

11) Line 349-350 “, labeling the cloud shadows has no significant ...”

12) Line 351 “We believe that when ...”

13) The experimental study is quite extensive and provides a satisfactory number of results. In general, I would like to see three additions:

• Could the authors include a third stage of dataset division, viz., apart from 1% and 3% cloud shadows percentage? The results are very interesting so, as a reader, I would like to see if the above findings constantly apply in a third level of cloud shadows balance, e.g., 5%.
• The discussion should be reorganized in a more dense format. I believe the reader could be more benefited if he/she can observe the advantages and disadvantages of different dataset divisions in a unified paragraph. As an example, a paragraph could discuss the performance and metrics obtained in the 1% of cloud shadows and a second one in the 3%. The last paragraph could summarize the advantages and disadvantages of each one of the above dataset divisions. Besides, a per-metric description of the results is already provided in Section 3.
• Are there any studies that can be included in order to compare your method? Given the novelty of the dataset, I suppose that the authors could only test, if possible, existing water body extraction algorithms on the existing data.

14) The level of English can be improved.

Overall, I really like the topic and the introduced approach. Hence, I will be more than happy to endorse acceptance of the manuscript, provided that the authors proceed with the above-mentioned corrections or justify possible arguments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

-Add related works section

-if it is possible compare the obtained outcomes with other studies

-add references where other methods are used to manage water resources

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have successfully replied to all of my comments. Thus, I am very happy to endorse the paper and recommend publication.