SAR Image Classification Using Markov Random Fields with Deep Learning

Round 1

Reviewer 1 Report

See attached file.

Comments for author File: Comments.pdf

Author Response

Dear reviewer 1, a point-to-point response to your comments updated, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

 

Comments to Editor:

 

The authors propose a classification algorithm for SAR images and combine superpixels and CNN into the framework of MRF. The conception of probability field is novel and reasonable, which optimizes the initialization of labels and the punishment of potential in random fields. Experiments on simulated and real images prove the property of the proposed algorithm, and the discussion on the number of superpixels demonstrates the robustness of the methods. The following are more detailed comments.

 

[1] Introduction ' Markov random field (MRF) has been widely used in both optical image and SAR image [3] ', the reference [3] only verifies the performance on the optical image, please cite more relevant papers to state this opinion.

[2] 'Fig.1. Flowchart of the proposed approach ' the image quality needs to be improved, and the division of fields should be clearer.

[3] In Section 2.1, the construction of RAG is based on which kinds of rules to link regions, please specify.

[4] In Section 2.2, the training process of CNN is not explained, please specify.

[5] Equations (11) and (12), the definition of Uprob is not mentioned, please rephrase.

[6] The text writing and reference format need to be modified, e.g., the title of tables states with a capital letter and abbreviation of the journal name.

 

 

Author Response

Dear reviewer 2, a point-to-point response to your comments is updated, please see the attachment,

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposed a Markov Random Field (MRF) based algorithm for SAR image classification. The initialization of region labels is obtained by CNN. Then, a probability field is constructed to improve the distribution of spatial relationships between adjacent super-pixels. Lastly, a novel region-level MRF is employed to classify the super-pixels. The method reduces the misclassification in pixel-level and rectify the misclassification in region-level. The experimental results of the algorithm on several datasets confirm the property of the method.

As a whole, the proposed way is feasible. However, some points should be further explained.

 

1. The feasibility of the manuscript innovation point should be considered.

 

2. How the initial classification results in pixel-level are transformed into classification results in region-level.

 

3. In the caption section of each figure it is advisable to briefly describe the content and function of its components.

 

4. There are many descriptive and grammatical errors in the manuscript. The descriptions of the figures and tables should be complete sentences, but there is no end-of-sentence symbol. There is an extra comma in figure 3 (line 128). The author is advised to check the paper properly for typographical and grammatical errors.

 

5. In the experimental section, all the methods of comparison are not referenced. Meanwhile, more recent methods (within three years) should be compared.

 

6. The evaluation metrics in the manuscript are only OA. From the data in the table, we can see that the classification effect is different for different categories and different datasets, whether more evaluation metrics can be used to evaluate each category separately and analyze the advantages and disadvantages at the same time.

 

7. The classification result map can't obviously show the classification effect of different methods, it is suggested that the authors use some way to be more obvious and finally give ground truth.

 

8. Whether the authors try to use different convolutional networks (different number of convolutional layers and different sizes) has an influence on the experimental results. Or using the current mainstream backbone network.

 

Author Response

Dear Reviewer 2, a point-to-point response to your comments updated, please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have answered all my questions satisfactorily and followed my suggestions.

Reviewer 2 Report

Although the author had revised the most of comments, there are still low-level errors in it which I had already pointed out clearly. I am sorry to feel that the author didn't take seriously with my opinions. Fourthermore, I suggested that some SOAT contributions should be investigated in your introduction.