Automatic Ship Detection Based on RetinaNet Using Multi-Resolution Gaofen-3 Imagery
Round 1
Reviewer 1 Report
comment to the authors are in the attached document: comments_to_authors.docx
Comments for author File: 
Comments.pdf
Author Response
Dear Section Managing Editor , 15 February 2019
Dear reviewers
Remote Sensing
Manuscript ID remotesensing-439733 entitled "Automatic Ship Detection based on RetinaNet Using Multi-resolution Gaofen-3 Imagery".
We really appreciate the positive feedback of the two referees and own many thanks to their reviews. We agree with these suggestions and have revised the manuscript accordingly. Below is our response to his/her comments resulting in a number of clarifications. We hope these revisions resolve the problems and uncertainties pointed out by the referees. In the manuscript and this file, the red and blue parts are revisions suggested by the two reviewers. The underline parts in the manuscript are those changed contents to improve the expressions.
Best Regards,
Chao Wang
Author Response File: 
Author Response.docx
Reviewer 2 Report
The work seems already passed through a first process of revision.
However, a large/moderate revisiting of English should further improve the clarity of concepts. Examples are in line 12 (multi-scales... what? properties,? characteristics?); line 33 (marine: better to say: ocean, sea); line 206-207 (the sentence is not clear and should be rewtitten); line 285 (score).
Other minor remarks:
- line 149: please specify what do you mean by hard examples? Are they ground truth?
- line 194 on using the focal loss as cost function. Perhaps this statement should be commented further. Has the use of focal loss as a cost function implication on the kind of result you generate or the result is general and quite invariant? Are they alternative?
A general comment: have ou tested the method in different sea conditions? I.e. using the ship detection algorithm also in different speckle levels?
Author Response
Dear Section Managing Editor , 15 February 2019
Dear reviewers
Remote Sensing
Manuscript ID remotesensing-439733 entitled "Automatic Ship Detection based on RetinaNet Using Multi-resolution Gaofen-3 Imagery".
We really appreciate the positive feedback of the two referees and own many thanks to their reviews. We agree with these suggestions and have revised the manuscript accordingly. Below is our response to his/her comments resulting in a number of clarifications. We hope these revisions resolve the problems and uncertainties pointed out by the referees. In the manuscript and this file, the red and blue parts are revisions suggested by the two reviewers. The underline parts in the manuscript are those changed contents to improve the expressions.
Best Regards,
Chao Wang
Author Response File: Author Response.docx
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.
Round 1
Reviewer 1 Report
This paper reports a machine-learning-based ship detection method for Gaofen-3 satellite data. The proposal itself is reasonable however; multiple modifications are required prior to be published.
1. In L. 68-71, the authors wrote that preceding researches “lack of large volume dataset to better evaluate the performance of object detectors on multi-scale” However, decreasing the resolution is not a difficult problem. One can cut the bandwidth or apply averaging filter to SAR images to simulate a worse resolution case. Adding (Gaussian) noise can simulate multiple sensor conditions. Therefore a better reason is required or the authors may simply note that they used Gaofen-3 experimentally. The authors can compare simulated resolutions (or in other words, 5-10m resolution images which are generated from 3m image) and actual worse images.
2. In L. 73-74, the authors wrote “To relieve this predicament, 9,974 SAR ship patch images cropped from SAR imagery are constructed and will be released in the future.” Please cite a specific reference for the dataset or describe from where, how the authors acquired the dataset and why the authors know that will be released in future.
3. In L. 182 and L. 193, the authors applied 10 times different values for the learning rate. Please briefly describe the reason.
4. In Fig. 5 and corresponding texts, what is the unit of “area” and “ratio”? Especially, what are the “height” and “width” of the bounding box for the ratio? Are they the size of range and azimuth pixel?
5. In L. 206-207, the authors wrote “with 16 bytes and are difficult for visual interpretation by human, they are stretched and restored with 8 bytes” Does this reduction affect results?
6. In L. 207-212, the authors wrote how they acquire the truth data. However, please describe it more precisely. Did the authors make it manually? Some researches use AIS (Automatic identification system) information for large ships and ignore small ones.
7. In Table 1, what are the width and height? Are they Range and Azimuth pixels? Please use specific terms.
8. In Section 3.2.1, please briefly describe or discuss why RetinaNet-3 showed the best.
9. In Section 3.2.2, the authors wrote "a large volume of dataset contributes" the accuracy. Contrarily, 5m and 8m resolution, the largest and the smallest, dataset marked the worst and the best score for MAP_1, respectively. The authors must prove their hypothesis further. At the same time, the authors must show the accuracy’s dependency on the volume of the dataset to prove their hypothesis.
10. In Section 4, in order to show the robustness of the method, it is better to discuss the dependency for the incidence angle, polarimetry and sea state too. As this research focus on the robustness or adaptability of the existing method for the ship detection, these discussions are more important for other researchers.
11. Minor modifications.
11-1. In L. 177, α_"t" -> αt
11-2. In L. 206, “Since” must be small characters.
11-3. In L. 227, “slighter higher” -> “slightly higher”
11-4. In L. 233, “Retinanet-3” -> RetinaNet-3
Author Response
Dear Assistant Editor,
Mr. Vladimir Maksimovic 7 November 2018
Remote Sensing
Manuscript ID remotesensing-380670 entitled "Automatic Ship Detection based on RetinaNet Using Multi-resolution Gaofen-3 Imagery".
We appreciate the thorough reviews provided by the referees. We agree with these suggestions and have revised the manuscript accordingly. Below is our response to their comments resulting in a number of clarifications. We hope these revisions resolve the problems and uncertainties pointed out by the referees. In the manuscript and this file, the red and blue parts are revisions suggested by two referees, respectively. The underline parts in the manuscript are the changed contents to improve the expressions.
Regards,
Chao Wang
Hong Zhang
Author Response File: Author Response.docx
Reviewer 2 Report
The authors applied the already published RetinaNet algorithm, which was developed for general object detection, to the more specific case of SAR ship detection. The algorithm’s performance is compared to the performance of other algorithms and shows highest mean average precision (MAP), for which the used focal loss is made responsible for. The paper could deserve publishing, as already some interesting results are shown. However, a major revision is required, as the paper is difficult to understand in its current form and some parts must be explained with more detail. Also some further analysis is required, in order to prove statements from the authors, which are currently not traceable.
Major:
Please let the paper be checked by an English proof reader. The paper contains multiple misspellings and misinterpreted/misinterpretable words, which impedes reading and understanding. At many places the text had to be interpreted based on context and educated guesses. It cannot be excluded that some parts were interpreted wrongly.
Page 5, line 165-168: More information about the manual inspection is required in the paper, as understanding the underlying dataset is especially important for every reader when deep learning techniques are used. Was each pixel labelled or rectangles drawn around the signatures? Which class labels were assigned (only ship and non-ship, or more). Was a human error accounted for (e.g. false labelling as ship classes of offshore platforms, ambiguities, sea signs, sea surface artefacts)? In the Introduction you mention the drawback of CFAR and feature based methods which are “predefined by human”, but now your whole dataset seems to be constructed by human expert knowledge. Or did you use ground truth data (e.g. AIS, campaigns…) to include samples where a human did not encounter the searched ship features (which are to some degree also predefined).
Figure 5/ page 6, line 210-215: The plots are not clear to me. Please provide more information about the cutting of sub images. What do you mean by ground truth bounding box? How can the chips have different height to width ration, when they are always 256x256 pixels? You use a sliding window to extract patches, does this mean each ship is in the dataset multiple times or does each patch contain a unique ship? I suggest to explicitly introduce what you understand under the terms: chip, patch, box, window and so on.
Section 3.2: Large ships should be detectable with nearly 100% by any ship detector, while small ships are difficult to distinguish from sea surface artefacts even by human experts. It should be described how many ships of different sizes were in the dataset (e.g. small ships <25m, medium ships <100m, and large ships >100m, you could also provide the ship size classes according to an official classification scheme, e.g. by EMSA). A performance comparison for different ship size classes could then be provided as an additional sub-subsection.
Table 4: This is an unexpected and interesting result, which deserves discussion in the discussion section.
Section 3.2.3: In the Abstract and Introduction you mention drawbacks of CFAR and feature based methods, but now you only compare deep learning based methods. Please motivate the three classifiers selected for comparison.
Figure 7: a) I recognize bright scattering at the lower right corner. Why is this not considered?
Section 4: I suggest to more discuss your experimental results. See also my comment for Table 4.
Minor:
Page 1, line 15: Your statement is very general and therefore not true. What does “making ship detection difficult” mean? Please be more specific. See also my comment below (Page 1, line 36-38)
Page 1, line 36-38: Both methods have been proven suitable for actual scenes, which is also shown in the reference [9] you give. Also generally stating that all versions of these both methods are useless on heterogenous SAR is a strong statement, which is not supported by your reference.
Page 2, line 69-71: RADARSAT-2 has more modes
Figure 1: Please increase readability/size of all text
Some possible misunderstandings due to your English skills, I stopped after the first page:
Page 1, line 14: sea clutter, not cluster, the error occurs at several positions in the paper
Page 1, line 33-34: Why is it imperative to process a large volume of data in order to do near-real time reactions?
Author Response
Dear Assistant Editor,
Mr. Vladimir Maksimovic 7 November 2018
Remote Sensing
Manuscript ID remotesensing-380670 entitled "Automatic Ship Detection based on RetinaNet Using Multi-resolution Gaofen-3 Imagery".
We appreciate the thorough reviews provided by the referees. We agree with these suggestions and have revised the manuscript accordingly. Below is our response to their comments resulting in a number of clarifications. We hope these revisions resolve the problems and uncertainties pointed out by the referees. In the manuscript and this file, the red and blue parts are revisions suggested by two referees, respectively. The underline parts in the manuscript are those changed contents to improve the expressions.
Regards,
Chao Wang
Hong Zhang
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The revised paper seems fine to be published.
Author Response
We appreciate the thorough reviews provided by the referees and handling editor. We agree with these suggestions and have revised the manuscript accordingly.
Reviewer 2 Report
comment to the authors are in the attached document: comments_to_authors.docx
Comments for author File: Comments.pdf
Author Response
Dear Assistant Editor,
Mr. Vladimir Maksimovic 4 December 2018
Remote Sensing
Manuscript ID remotesensing-380670 entitled "Automatic Ship Detection based on RetinaNet Using Multi-resolution Gaofen-3 Imagery".
We really appreciate the positive feedback of the two referees and own many thanks to their reviews. We agree with these suggestions and have revised the manuscript accordingly. Below is our response to his/her comments resulting in some clarification. We hope these revisions resolve the problems and uncertainties pointed out by the referee. In the manuscript and this file, the red parts are revisions suggested by the referee. The underline parts in the manuscript are those changed contents to improve the expressions.
Regards,
Chao Wang
Author Response File: Author Response.docx
