Automatic Ship Detection Using PolSAR Imagery and the Double Scatterer Model
Round 1
Reviewer 1 Report
Please mention the performance of your approach in cases such as small islets and ships near the coast. Concerning the conclusions sections: The training on very high-resolution data in C-band and the application on high resolution data in L-band shows that the method is independent of data resolution and SAR band. If so, it should be mentioned.
Detailed comments:
Page 2, Last paragraph: In order for the read to quickly comprehens experiment please write down that the 1st and 2nd sets used for training and the 3rd as testing.
Page 8, Line 269: How this decision is taken automatically.
Figure 7: The far left ‘contribution of trihedral’ circle is redundant. It is clearer to have one instead of two circles and discriminate three cases underneath. Unless if there is a special reason to leave both, which I don’t understand. IF so please be more explanatory
Figure8 : Add legend for the left: colorscale i.e. grayscale and right :class 1 (white) as target and class (black) 2 as clutter
Page 11, Line 315: What FoM stands for?
Reference 22 is incomplete
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Please see the attached PDF file.
Comments for author File: Comments.pdf
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper presents a ship detection method for PolSAR imagery that is largely based on the authors prior work on developing the Double Scatterer Model. The introduction reads fine and includes a good number of the relevant literature, with the following sections describing prior art related to PolSAR scattering and the utilised model also well presented (though I'm not sure there is much need for the last column of Table 1, it's obvious what z is in each case).
However, there is room for improvement in the presentation of the proposed decision tree algorithm in Section 4.
You mention you train a supervised algorithm to separate targets from clutter. What is the algorithm learning exactly? Is it the contribution rates that set the decision thresholds you mention later on? It's not clear at this stage. Also what data is it trained on? You describe 3 image products earlier on, are these the entirety of your test/train datasets? How do you know training ends? With what criteria does it do so? There is very little information here for someone who would try to replicate your algorithm.
Further on, you mention some contribution rate thresholds. Clarify where those came from - are they learned from the algorithm? Are these data specific? Are they the same for any image given to the algorithm? Do they need to be accurate to the 9th decimal? Also, the value you quote in the main text and the value shown on the graph are different. First of all check these are correct, secondly perhaps it's better to label this on the graph as a threshold value rather than a long number which may or may not be right.
Also, the left hand side of your tree algorithm - you check the trihedral contribution and if it is under a threshold you move to the new stage and in that new stage, you check the same thing again but with a lower threshold now? It's unclear what the difference is? Why not immediately check for the lower value of 0.765% and eliminate as clutter anything above it? It's really unclear what this stage is supposed to be/do - you simply say a "more in-depth analysis is needed because the entropy/uncertainty of the decision remains high", which tells me next to nothing about what that analysis is and how it works.
You really need to rework this section because frankly it would be impossible for someone to fully understand, let alone recreate your work based on the information provided here. There is too much missing in terms of implementation detail, training regime, parameter tuning and all sorts of detail.
Finally, moving on to your results section. You demonstrate performance on a single image product, against a single other algorithm. And if anything, that is the built-in CFAR in SNAP. While I think including this method is a good idea to provide a general benchmark, it is not a state of the art method. It is also not a PolSAR specific method but rather a simple single-channel method.
I would suggest you significantly rework your experimental section. You should include at least a state-of-the-art method against which you compare your results, and preferably this should be a method specifically developed for PolSAR data, since this is your focus. And at the very least quote results on all of your available images. A single example compared against simple CFAR would frankly not be sufficient experimental validation for a conference abstract let alone a full-length publication.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Only part of the corrections are perfromed, however manuscript is acceptable.
Reviewer 3 Report
I welcome the authors response, and in particular their clarification and restructuring of certain sections as discussed - the method is much clearer to understand as a result.
I would have liked to see more experimental comparisons but if no other methods are available I guess this will have to do.
