Dark Spot Detection from SAR Images Based on Superpixel Deeper Graph Convolutional Network

Round 1

Reviewer 1 Report

Oil spills are a serious environmental problem. The authors work in the wrigt direction to make our world better. I like the paper and material presentation and have only a few comments.

 

1. There are some interesting studies in oil spill detection that could be mentioned in the deep learning section of the Introduction. For example:  

Kan Zeng, Yixiao Wang, A Deep Convolutional Neural Network for Oil Spill Detection from Spaceborne SAR Images

Yekeen, S.; Balogun, A. Automated Marine Oil Spill Detection Using Deep Learning Instance Segmentation Model.

Abdul BasitMuhammad, Adnan SiddiqueMuhammad Adnan Siddique, Muhammad Khurram Bhatti,  Saquib Sarfraz, Comparison of CNNs and Vision Transformers-Based Hybrid Models Using Gradient Profile Loss for Classification of Oil Spills in SAR Images

etc.

The authors should check the latest related research in their area.

 

2. It would be good to rebuild Figure 2 to be more simple and clear. Figure 3 is a good example.

 

3. Figure 5 could be split into two figures that could be placed near the corresponding parts of the paper. It is hard to jump over text and the figure. 

 

4. I recommend placing Figure 6 on page 12 even if it will require removing the bottom line from the figure.

 

5. There are no discussions in the Results and discussions section. Some ideas and directions are presented in the Conclusion section. It would be good slightly reorganize that part and also add some limitations of the method if there are any.  

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

This study mainly focuses on the work of detecting oil slick on the sea surface based on SAR, and proposes a method of SAR image dark spot detection based on super pixel depth map convolution network, which is helpful for improving the accuracy of related targets. In addition, the method is applied to detect oil slick black spots in six typical large-scale SAR images covering the Baltic Sea, and good results are obtained, which proves the reliability and effectiveness of the method. Compared with the previous work, the model in this study is more robust and effective. The specific suggestions are as follows: (1) The results section and the discussion section should not be put together. The current content is very difficult to read. It is suggested that the structure of this manuscript be adjusted for this problem; (2) The location of the image in Figure 5-7 should be shown on the plan; (3) To sum up, a moderate revision is recommended.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

The aim of the study described in this paper is to detect dark spots on the ocean's surface for oil spill monitoring. The method developed, tested, and compared is in three parts: 1) the segmentation of SAR images into "superpixels"; 2) the selection of the most relevant features for these; and 3) the classification of the "superpixels" for dark spot mapping. The explanations concerning the methodology could be a bit more detailed. The results are however interesting and compared to several other methods. The presentation is very clear and to the point. I am not able to judge the originality of the content.

General comments
1. The transition between the detection of dark spots and the more specific detection of oil spills could be clarified.
2. The captions of the figures and tables could be more comprehensive in order to be self-sufficient.

Specific comments
1. [23] Acronym "SVM-RFE" not previously defined.
2. [24] "Optimal" according to which criterion?
4. [72] I would suggest not to describe ML algorithms as "intelligent".
5. [107] You might not have meant to write that images can detect dark spots. Maybe dark spots can be detected in images rather?
6. [113-119] This paragraph just repeats what has been mentioned in the paragraph above.
7. [142: Table 1] Why list the main characteristics of all available modes while only one was used in the study? I would suggest replacing the table by a reference.
8. [218] Edge features are described as "he" here but as "h_e" in Eq. (1). Same comment for N(v) and N_(v).
9. [Eq. (1)] The definition of m is missing.
10. [Eq. (4)] If the graph is undirected and u belongs to N(v), what is the purpose of the indicator function here?
11. [237-244] This paragraph is not clear. I think it is worth rewriting and expanding.
12. [255] BASS not previously defined.
13. [264] What is t?
14. [268] Overall Accuracy is not well suited for unbalanced datasets. I suggest using F1-score or Matthew's Correlation Coefficient.
15. [276] How come there are 137 feature values here but 48 in Section 3.2?
16. [280] You mention using F1-score to select features, but what are you classifying exactly here? I guess oil spills vs lookalikes among dark spots, but this is not clear at all from this section or from Section 3.2!
17. [292] Percentages were inverted in relation with the 13 and 16 features mentioned in 283-284.
18. [Fig. 4] The legend is missing.
19. [Table 3] I would recommend not to describe the subset of features as powerful.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Accept.