A Gamma-Log Net for Oil Spill Detection in Inhomogeneous SAR Images
Round 1
Reviewer 1 Report
This study is about a method to detect oil by applying Gamma correction to SAR images of the sea surface and segmenting the images.
I believe that this method is promising and significant.
However, it should be written considering that there are many readers of Remote Sensing who are not familiar with artificial intelligence.
In neural networks, the optimal parameters would been obtained based on training data, and the parameters should have been validated by applying them to data that was not used for training.
However, methods such as UNet used in this study do not describe the tuning of neural network parameters.
Does this mean that the optimal values were obtained only for the Gamma correction parameters?
Or, it is not clear whether other parameters are also obtained together in the learning process.
Figure 4 schematically depicts the procedure of the method used in this study, but it is difficult to understand what kind of procedure was used.
First of all, by feature map, do you mean the original input image?
There is equation 5 for global mean pooling (what is C?), What are the other 1*1 and 3*3 convolutions?
What kind of convolutions are applied to the input pixel values?
What is the next 2* sigmoid?
What sigmoid function is being applied to the processed pixel values?
What operation is responsible for the three images (1*1, 3*3, global average pooling) becoming one again (green image)?
I have pointed out only a very small part of the problem, but the explanation in this paper is not sufficient for the reader of Remote sensing.
Author Response
Thank all the editors and reviewers for the efforts of processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have carefully revised the paper according to all the suggestions. All the changes have been marked in red in the revised manuscript. Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper proposes a new algorithm for detecting oil spills in inhomogeneous SAR images. The idea is to improve an image segmentation architecture for this task using the Gamma-Log module which allows adaptive adjustment of SAR images. Experiments on real-world data are provided using six common network architectures to demonstrate that the proposed method works effectively and compares favorably against the state-of-the-art. The paper studies an important problem and despite its simplicity, the Gamma-Log module is effective. I like the paper and have the following comments to be addressed before publication. 1. Please add more motivation behind using the Gamma distribution and its advantage over using other distributinal form to give the reader a better perspective about this chocie. 2. In Tables 2 and 3, please perform experiments several times and include the standard deviation? The numbers are close and we need standard deviations to conclude that the differences are statistically meaningful. 3. Could you perform experiments by varying the value of gamma_i? I would like to study the effect of the value for this parameters on performance. 4. An idea to improve SAR segmentation is to benefit from domain adaptation by first training a network on optical images. This idea has been used successfully to improve image segmentation in different imaging modalities. Please include a discussion on this topic and cite the following recent works: a. Farahani, M. and Mohammadzadeh, A., 2020. Domain adaptation for unsupervised change detection of multisensor multitemporal remote-sensing images. International Journal of Remote Sensing, 41(10), pp.3902-3923. b. Stan, S. and Rostami, M., 2021, May. Unsupervised model adaptation for continual semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 3, pp. 2593-2601). c. Zhang, C., Feng, Y., Hu, L., Tapete, D., Pan, L., Liang, Z., Cigna, F. and Yue, P., 2022. A domain adaptation neural network for change detection with heterogeneous optical and SAR remote sensing images. International Journal of Applied Earth Observation and Geoinformation, 109, p.102769. d. Liu, Z., Zhu, Z., Zheng, S., Liu, Y., Zhou, J. and Zhao, Y., 2022. Margin preserving self-paced contrastive learning towards domain adaptation for medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 26(2), pp.638-647. 5. Is it possible to release the code for the paper on GitHub?Author Response
Thank all the editors and reviewers for the efforts of processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have carefully revised the paper according to all the suggestions. All the changes have been marked in red in the revised manuscript. Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper proposes a Gamma-Log Net architecture to adjust the inhomogeneous SAR images which refer to low contrast and difficult to distinguish between background and targets.
According to figures 6 and 7, the proposed Gamma-Log module works even worse than the non Gamma-Log module.
Another issue is that, it's mentioned in conclusion that marine natural phenomena can suppress or enhance the reflection of SAR echo signal. However, the effect of marine natural phenomena is not shown in any part of the manucsript.
Author Response
Thank all the editors and reviewers for the efforts of processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have carefully revised the paper according to all the suggestions. All the changes have been marked in red in the revised manuscript. Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors have addressed all my concerns.
Author Response
Thank all the editors and reviewers for the efforts of processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. Thank you again for your suggestion.
Reviewer 3 Report
Again, I think the proposed Gamma-Log module works even worse than the non Gamma-Log module, according to figures 6 and 7. Please check the region in the blue circle as shown in the attached file. It's better to show the model predicted image substracted by the ground truth and to compute the area of differences between the model predicted image and the ground truth.
Comments for author File: Comments.pdf
Author Response
Thank all the editors and reviewers for the efforts of processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have carefully revised the paper according to all the suggestions. All the changes have been marked in blue in the revised manuscript. Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 3 Report
The region in the purple circle in the revised figure 8 is quite different to that in the figure 6 for the UNet++. Also, I don't think the Gamma-Log module works better for the Attention-UNet,
Comments for author File: Comments.pdf
Author Response
We would like to thank all the editors and reviewers for their efforts in processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have carefully revised the paper according to all the suggestions. To distinguish from the first revision, all the changes have been marked in orange in the third revised manuscript.
Please see the attachment.
Author Response File: Author Response.pdf
Round 4
Reviewer 3 Report
Please upload your data and codes. It's better that the editor of remote sensing to run the code with the data to check the results.
Author Response
We would like to thank all the editors and reviewers for their efforts in processing our manuscript. All the comments provided are very helpful for us to improve the paper’s quality. We have uploaded code and data to Github. Please see the attachment.
Author Response File: Author Response.pdf
