TCSPANet: Two-Staged Contrastive Learning and Sub-Patch Attention Based Network for PolSAR Image Classification

Round 1

Reviewer 1 Report

This manuscript introduces an unsupervised PolSAR image classification model.  The advantages of using unsupervised learning vs. supervised learning are explained (less human effort in manually classifying pixels).  The results are comparable to those from supervised methods, though sometimes worse. 

There were a few points I feel should be addressed before the manuscript may be considered for publication:

1. It is unclear how or why certain parameters were selected the way they were.  For example, in Equation 1, a threshold of 4 is used as a lower bound for the size of ST divided by the size of T.  Some more explanation is needed as to why this number is 4, and not some other number, say 5 or 6.
2. The procedure on lines 203 and 204 depends on the ordering of the pixels since only the upper-triangular part of the matrix is calculated.  Instead, it would make sense make the matrix symmetric by setting it equal to (A+A^T)/2
3. In Equation (4), the ordering of patch clusters matters.  Some comment should be provided as to the effect of different orderings and how it might affect the final result.
4. The caption of Figure 8 has a  typo: sprcific should read "specific"
5. How is gamma chosen in Equation 11?
6. The reference on line 407 for CV-CNN is not given (just a question mark)
7. Lines 504 and 522 in Appendix A - learining should read "learning"

Author Response

Please see the attachment. Thank you.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents a technique for patch based PoISAR image classification. The technique is named TCSPANet, which integrates two stage contrastive learning and sub-path attention mechanism. Experiments are performed on three PoISAR image datasets and detailed results and analysis are presented.

The following comments should be addressed:

• At times, the writing is unnecessarily very lengthy, which can be cut short throughout the paper. For example, justification for patch based approach in Line 57, (even the entire page) is sort of well known. A short narrative with appropriate references will do. For pictorial examples, please try to use one figure for multiple things to cover instead of separate images for each small/detailed explanations. Finally, the writing is verbose. One-third of the manuscript should be cut down.
• There are several references missing, [?] in line 407. There are a few more. Care should be taken to prepare and check a manuscript.
• There are many unjustified thresholds used, e.g., 4 in Eq 1 and Scal value 2.
• Link equation numbers in Algorithms in Appendix.
• Other methods eg Random Forest and Boosting (XGBoost), and state of the art deep learning models that are used traditional RGB image classification/satellite image classification, should be tested and discussed.
• In Tables, the separate boldface for standard deviation numbers look odd. I am not sure is it a norm.
• The datasets may have more classes (detailed classes), e.g., Flevoland has crop specific classification labels. It will be interesting to see the results.
• Confusion matrix should be added.
• Statistical significance test of the results should be included.
• Computational time for training and inference should also be reported.

Author Response

Please see the attachment. Thank you.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors addressed my comments reasonably well. The paper can be accepted.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.