Assessment of Machine Learning Techniques for Oil Rig Classification in C-Band SAR Images

Round 1

Reviewer 1 Report

Based on the attribute extraction capability of classic convolutional neural networks and the classification capabilities of various classic machine learning algorithms, this article studies the oil rig classification and evaluates the performance of different methods. The paper does not explain the reason for the parameter setting of the used machine learning method. For the experimental results, it seems to lack result analysis. Some unclear descriptions do affect the readability of the manuscript, and some comments should be considered for improvement before recommending publication.

 

1) Regarding feature extraction, why did the author use VGG16 and VGG19? What is the difference between the two? Also, have the authors considered using a more advanced network model? If convenient, please provide comparative experimental results of other network models.

 

2) For Table 3, What is the basis for the setting of these parameters? Is it empirical or based on references?

 

3) The authors should give some analysis of Appendix A and not be limited to giving numerical results.

 

4) It is recommended to check the language problem.

Page 11, line 303.

The classifiers were evaluated considering six different methods (M1, M2, M3, M4, M5).

‘six’ or ‘five’?

Author Response

Dear Reviewer,
Thank you very much for reading the article. Your suggestions will greatly contribute to the development of my academic work.
Please "see the attachment" for the answers to your recommendations.

Thank you very much

Fabiano Gabriel da Silva

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Although the topic is interesting, the manuscript has several flaws in technical aspects and writing. The manuscript presented such a report, not as an academic paper. The manuscript should be redesigned in methodology and improve the writing and technical parts to present the result more readable. I regret to say in this stage it is not sufficient for publication.  

Author Response

Dear Reviewer,
Thank you very much for reading the article. Your suggestions will greatly contribute to the development of my academic work.
Please "see the attachment" for the answers to your recommendations.

Thank you very much

Fabiano Gabriel da Silva

Author Response File: Author Response.pdf

Reviewer 3 Report

good paper  - small suggestions noted in the attache

Comments for author File: Comments.pdf

Author Response

Dear Reviewer,
Thank you very much for reading the article. Your suggestions will greatly contribute to the development of my academic work.
Please "see the attachment" for the answers to your recommendations.

Best regards,

Fabiano Gabriel da Silva

Author Response File: Author Response.pdf

Reviewer 4 Report

The manuscript "Oil Rig Classification Based on Artificial Intelligence Using C-Band SAR Images" presents comparisons among different classification methods in a collection of Sentinel-1 SAR images transformed by a classical convolutional neural network architecture (VGG-16 and VGG-19).

The text makes a distinction between machine learning (ML) and deep learning (DL) that is misleading: First, DL is a specific technique inside ML field; Second, the paper extracts features (embeddings) using deep neural network architectures, and by definition is part of the DL toolset.

It is also hard to understand the contribution of the manuscript, as it replicates the results from manuscript reference [26], and the additional proposed methods are variations of "parameter search" that have little novelty.

 

 

Please consider the following comments and questions:

 

1. The manuscript title is too broad, with the term "Artificial Intelligence" not being justified. Please consider a title that better describe the methods explored in the text.

 

2. From "data set contains 400 images (patches) in VH and VV polarization" (page 3, lines 111-112), it is not clear how the dataset is formed. Please clarify in the text:

2.1. How unique many SAR images were used?

2.2. What is the time and condition of those images?

2.3. How the targets were detected and extracted from the images?

2.4. What data was used to determine ground-truth labels?

2.5. What are the sizes of the 400 image patches? How they relate to the different target sizes and the image resolution?

 

3. The dataset suggests a total of 400 patches equally balanced between "platform" and "ship" classes. Are all those targets distinct? Are they multiple realization of smaller number of platforms/ships? How many unique platforms are present in the dataset?

 

4. The text should be clear how the test dataset is formed. How many unique targets in the test dataset are also present in the training dataset (different patches, same target)?

 

5. From "after extracting attributes with the CNNs, four different data sets are created" (page 7, line 201). Please clarify: How the CNN model was trained, or if pre-trainined weighs were used, how they were trained and what transformations were applied to the SAR dataset to match the original network data training distribution?

 

6. What is the size of the extracted embedding? The size of 4096 is the output of the last dense layer before the prediction layer of the VGG-16 (or VGG-19) network. This means that you are using CNN layers + 2 dense neural network layers of the original architecture?

 

7. From "the only study available in the literature for maritime target classification in Sentinel-1 SAR data based on ML techniques" (page 7, lines 215-216). 

There are many papers available using Sentinel-1 SAR data, for example every publication that uses the OpenSARShip dataset. I believe the "ML techniques" is misleading, specially when a CNN model is used to extract features.

 

8. Table 6 presents the results as the average accuracy over the 50 bootstrapped groups? I believe the "3 decimal digits" as presented in the table is not consistent with the number of bootstrapped experiments. Please include confidence intervals to all predictions, to help readers properly interpret the results.

 

Author Response

Dear Reviewer,
Thank you very much for reading the article. Your suggestions will greatly contribute to the development of my academic work.
Please "see the attachment" for the answers to your recommendations.

Best regards,

Fabiano Gabriel da Silva

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Thanks to the author for preparing the cover letter and inserting the new materials in the text.  the manuscript seems improved a lot and is more readable, but still, there are some questions and commenst.

-  Still, the main input is not clear, and is it the main contribution is the comparison between some methods for the classification of a marine target?

- Could you present the image processing result of different methods of target detection by AI methods?

- Which polarization mode showed the better result in the classification of the target? VH or VV? 

- What application did you apply for IW mode? did you use phase data or just intensity? 

- Why convert the amplitude to zero-sigma, not beta nor gamma? 

- How do you use the GRD and IW mode together? is it included in 400 scenes of the image? either for each scene, do you download two modes of IW and GRD? if so why? 

- Did you have any field observation and accuracy assessment of results besides statistical techniques? 

By presenting of results of different classifications, images can be more evidence for the reader to see how the methods worked. especially the accuracy assessment is very important. 

Author Response

Dear Reviewer,

Thank you very much for reading the manuscript and for all the suggestions made.
"Please see the attachment" which contains all the responses to your remarks.

Yours sincerely

Fabiano Gabriel da Silva

Author Response File: Author Response.pdf