Research on Input Schemes for Polarimetric SAR Classification Using Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors introduce the research on the input scheme with more parameters compared to the conventional number of parameters for Polarimetry-SAR(PolSAR). And also, you can show the effective scheme and the improved results using the deep learning process. It is interesting and helpful for prospective readers who are studying the SAR image processing.

However, to publish it in this journal, I recommend some modifications and improvements.

1. I think that T-Matrix is one of the important keys in this proposed technique. Although I can understand and guess the meaning of the T-matrix, this manuscript is insufficient to understand the process by which this matrix is derived. Please modify and add the additional explanation.

2. I think that the location of Section 3.2 in the structure of the paper is ambiguous in the flow of the research. 

3. In section 3.3, according to schemes, parameters are used for the polarization features. For the prospective readers, please explain the reasons why the author chose the parameters in each scheme.

4. In section 3.4, the authors mention 'The network utilizes the cross-entropy loss function, as expressed in Formula (7)'. I think the cross-entropy loss function is 'Formula (8)', not (7). And, please explain more clearly how to apply this function in your method. 

5. As mentioned in section 3.4 and figure (1), the input data size is defined as 64 x 64 x n. Does it mean the number of azimuth bins(64) and range bins(64) with polarization features? In the data explanation for verification, the size is 7882(range) x 9072 or 9070(azimuth). Please add the information.

Based on the experimental results, we can understand the improved results. It is meaningful.

Comments on the Quality of English Language

The English language is written in an easy-to-read manner for relevant researchers.

Author Response

The authors, above all, would like to thank you for your comments to help to improve the manuscript. All the comments are seriously considered and the manuscript is refined correspondingly. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions. Based on the instructions provided in your letter, we have submitted the revised manuscript with revisions highlighted in a different color (red).

Please refer to the attachment for details

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Feature selection is meaningful, However, this manuscript only analyze several polarimetric features, and deep models will reduce performance variations with minor feature differences. So, I do not think changing one or two features can have obvious classification variations with deep models since deep models itself is an uncertain model. The detailed comments are given as follows.

1. More multi-feature based methods should be cited, and some other polarimetric and spatial features should be discussed.

2. Generally, deep models are end-to-end classification, with input source data information including the 9-dimension column vector or complex-valued information. However, this manuscript use scattering features cannot achieve the end-to-end classification. 

3. The multiple features are incomplete, some other scattering features, spatial and textual features should also be considered during discussion.

4. In general, feature selection is essential for classification. How to select features and what is the principle or criterion should be the main research context.

5. The evaluation metrics are not enough. More evaluation metrics should be added.

6. In addition, the incomplete features with 6-parameter obviously cannot obtain good performance. At least, the feature should be the original data with 9-dimention. Then, you can add other features to verify their complementary.

7. Experiments should be conducted multiple times, and statistical accuracies should be given to verify its robustness.

8. I think you should focus on using amount of features, and how to select valuable features. Rather than analyzing the incomplete features to reduce the performance.

9. More datasets should be added, and other network should be compared, ablation study should be added to verify which is the main contribution:The features or the network?

Comments on the Quality of English Language

The English is well.

Author Response

The authors, above all, would like to thank you for your comments to help to improve the manuscript. All the comments are seriously considered and the manuscript is refined correspondingly. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions. Based on the instructions provided in your letter, we have submitted the revised manuscript with revisions highlighted in a different color (red).

Please refer to the attachment for details

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

General Comment:

Research on input schemes for polarimetric SAR classification using deep learning in this study using the PolSAR images of Gaofen-3. The research method is reasonable and of great significance to improve the application of polarimetric SAR satellite images. Therefore, I suggest this paper could be accepted for publication after Major Revision. The detailed suggestions are available as follows:

 

Specific comments:

Comment 1: In the Introduction, there is not enough information provided about the current SAR classification algorithms based on CNN in the introduction.  Additionally, a comprehensive review of a related topic is necessary. State-of-the-art polarization scattering features methods should be added.

Comment 2: Schemes 1-7 are proposed to increase the accuracy of classification using some polarimetric features. What’s the difference between them? An explanation should be provided if there is a special reason for this.

Comment 3: Why is the incidence angle of experiment images in the range of 30 degrees to 40 degrees? An explanation should be provided among them.

Comment 4: Both the AlexNet and VGG16 are not new CNNs, have you ever tried to use the state of art networks to validate the conclusion?

Comment 5: An explanation should be given of the different normalization methods for polarimetric features.

Comment 6: There should be more details bout the quantitative comparison in experimental results.

Author Response

The authors, above all, would like to thank you for your comments to help to improve the manuscript. All the comments are seriously considered and the manuscript is refined correspondingly. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions. Based on the instructions provided in your letter, we have submitted the revised manuscript with revisions highlighted in a different color (red).

Please refer to the attachment for details

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors addressed most of my concerns.

I think it would be good to explain one more thing. I asked what ‘64’ meant. I'm confused with other data of size 7882 (range) x 9070 (azimuth).

Author Response

The authors, above all, would like to thank you for your comments to help to improve the manuscript. All the comments are seriously considered and the manuscript is refined correspondingly. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions. Based on the instructions provided in your letter, we have submitted the revised manuscript with revisions highlighted in a different color (blue).

Please refer to the attachment for details.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript has been improved greatly. However, some questions should be discussed.

 

1. Can you show the experimental results of different schemes without network? Such as features connecting with only a simple classifier.

2. Some multi-feature and deep learning related references should be added, such as

[1]Complex matrix and multi-feature collaborative learning for
polarimetric SAR image classification  [2] Polarimetric Multipath Convolutional Neural Network for PolSAR Image Classification [3]Spatial feature-based convolutional neural network for PolSAR image classification

Author Response

The authors, above all, would like to thank you for your comments to help to improve the manuscript. All the comments are seriously considered and the manuscript is refined correspondingly. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions. Based on the instructions provided in your letter, we have submitted the revised manuscript with revisions highlighted in a different color (blue).

Please refer to the attachment for details

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed the issues well. I recommend to publish in current forms.

Author Response

    The authors, above all, would like to thank you for your comments to help to improve the manuscript. We thank you for taking the time to review the manuscript and appreciate all your comments and suggestions.

Author Response File: Author Response.docx