Review Reports

Round 1

Reviewer 1 Report

The paper presents the classification of sea ice based on quad-pol and hybrid-compact-pol data acquired by the Gaofeng SAR sensor in C-band.
Hybrid-pol data are simulated from the quad-pol data. For training and verification Landsat data are used.

The adopted approach includes first the identification of suitable co-located radar and optical images. The separability index is used to reduce the number of polarimetric features from 70 to 19. Finally, groups of features are identified to perform three different classifications prior to their verification. The obtained accuracies are remarkable. However, the results of the comparison of quad-pol and hyberid-pol results is not as I would have expected. I have several important comments:

1. When selecting the features/parameters for evaluating the (simulated) hybrid polarimetric mode you did not take into account the most realistic HW implementation of hybrid-compact polarimetry,
i.e. right (or left) circular polarisation on transmit and simultaneous linear H and V polarisation on receive (as operated by RISAT-1, ALOS-2 and RCM). In this case circular polarisation on receive is not feasible and some of the parameters will not be available. I suggest replacing the selection of parameters in group 2 (see. Fig. 8 and paragraph below on page 13) in a way to take this into account. e.g. you could select \sigma^0_rh instead of \sigma^0_rl. You also need to comment on the motivation for this selection.

2. You need to add additional information on the configuration of your SVM classifier. Which SVM kernel did you use for distance computations? And which are the hyper parameters? It is common knowledge that SVM is not an optimum classifier for multiclass problems. So please add some more words on motivation, why SVM has been chosen. What would have been the alternatives? (section 4.3.)

3. Second comment on the SVM: It is not clear, why the SVM has been trained just one single time using all 7 features (group 1 and group 2 parametes, see section 4.3, pg. 13, line 424). I would have expected the training of the classifier independently for each group. In reality, you will not have the full amount of quad-pol parameters to train sets of hybrid-pol data.
Isn't this a source of error? Please explain! In my opinion, this could explain some discrepancies/doubts in the reported classification results.
My strong recommendation is to redo the evaluation with independent training!!

4. Depending on the outcome of the recommendation in 3, the first paragraph of the discussion section (pg. 17) needs to be re-written.
In my opinion, classification using hybrid-pol parameters should be as good or inferior to classification results using quad-pol parameters.
There is no theoretical reason, why hybrid-pol parameters should give addional classification improvements compared to pure quad-pol results.     
My fear is that you derive wrong conclusions because of improper methodology using the SVM classifier (see comment 3).
Also with respect to Chandrayaan-2, you might have a wrong understanding on the motivation for using both, quad-pol and hybrid-pol, modes on that orbiter (pg. 17, line 505-509). Please revise!

Minor comments:
- incomplete sentence: pg. 2, line 51: "To obtain high-resolution sea ice classification information" can be removed.
- typo: pg. 6, line 193: col-pol -> co-pol
- pg. 6: duplication of first two paragraphs in section 3
- typo: pg. 8, line 300: in in Table 1
- grammar pg. 9, line 320: possible candidates to be classified. -> possible candidates to be used in classification. (You do not classify the candidates!!)
- missing verb: pg. 9, line 331: S5 <<provides>> the best match
- Table 2: What is the size in [m] of one pixel?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Synthetic aperture radar (SAR) has been used to monitor the condition of sea ice, but there is limited work on the classification of various sub types of sea ice. This paper proposes using a "separability index" to evaluate the effectiveness of 70 parameters for sea ice classification. Then, based on the discovered effective parameters, SVM machine is used for sea ice classification, and the classification results are evaluated based on manually interpreted ice maps derived from image data. The paper presents a detailed derivation process, and the final classification results show high accuracy, with sufficient discussion of the experimental results. However, more background information can contextualize the research and help readers understand the motivation and significance of the study.

The specific comments are as follows:

1. The paper does not introduce the differences between new ice, young ice, and first-year ice and their scientific significance in the introduction section. Additionally, the paper does not describe the characteristics and differences of these classifications in the description of Figure 4.

2. Are there any other works on classification of sea ice subtypes, and how does this paper compare to them in terms of innovation and advantages?

3. What are the limitations of using high-resolution optical image data for sea ice classification? Is it mainly due to the lack of nighttime data?

4. This paper mentions to identify which parameters are the most effective for different regions and under various seasonal or environmental conditions. However, the paper does not use data from different seasons. What impact does this have on the results?

5. Has the paper considered reducing the redundancy of the 70 parameters? For example, performing Principal Component Analysis (PCA)?

Minor:

No full name of the abbreviation "OW" when it first appeared.

Grammar error, line 147, “The data in the red rectangle with about 12 km long and 147 9.6 km wide…”

Line 253: “polarimetric asymmetry (PA)，” wrong format of comma

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Please read the attachments.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I appreciate the replies and clarifications in the manuscript given by the authors. They have included the requested changes with respect to the chosen classifier (changed from SVM to random forest), the choice of hybrid-pol features and the discussion on Chandrayan mission.

I do not have further comments except for some abbreviations:

- OW, NI, YI and FYI shall be defined earlier in the abstract (line 16)
- CIS shall be introduced on line 42

Final remark: The issue about different classification accuracies in quad-pol and hybrid-pol sub-sets might also be due to the different number of features for the quad-pol and hybrid-pol cases (3 vs. 4).  But I understand, that this cannot be answered with the present data set because of the other uncertainties mentioned by the authors.