Multi-Scale Polar Object Detection Based on Computer Vision
Round 1
Reviewer 1 Report (Previous Reviewer 1)
To my surprise and regret, I did not find a response from the authors to my comments (Review #1) on the first version of the manuscript. This may be why, when the authors prepared the second version of the article, my comments were largely ignored. For this reason, I cannot recommend the second version of the article for publication. Particular comments are as follows.
1. The paper demonstrate ordinary routine work on application of very popular tools for objects detection and has not led to serious scientific results with SCIENTIFIC novelty that deserve to be published in such a reputable journal as Water.
2. The reviewer could not find in the text either the justification or a formal description of the changes that improve the conventional approaches.
3. The article suffers from the main disadvantage inherent in many other articles on neural networks: there is no strict formal justification for the proposed models; the quality of the proposed solutions is justified by a computational experiment, the results of which are difficult to verify.
4. Many figures are not provided with a sufficient and exhaustive explanation, such as Fig. 9; other figures seem to be redundant and does not explain much, for example, Fig. 12.
5. The text abounds in stylistic mistakes and typos, therefore, its English checking is highly recommended. For example, the metric "mAP" in many places is written as "map"; "Ture" (page 157, line 166) must be "True", etc.
6. The manuscript, in its present form, cannot be recommended for publication in Water.
The new version has new mistakes. See the comments above.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report (Previous Reviewer 3)
The manuscript has been thoroughly revised, and the authors have done commendable research. The early detection of sea ice will help to avert ice-collision of the navigating ships.
All the suggestions have been incorporated so the manuscript can be accepted in its present form without further revisions.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report (New Reviewer)
In this paper, the authors present an enhanced YOLOv5 model that incorporates the squeeze-and-excitation networks (SE) attention mechanism, the fast spatial pyramid pooling (SPPF), and the FReLU activation function. This method is tailored to detect five distinct types of sea ice images. Its performance was benchmarked against several renowned methods, including YOLOv3, YOLOv4-tiny, Faster-RCNN, and the original YOLOv5. While the paper is articulate and provides clear insights into the research and comparative experiments, it could benefit from a more detailed explanation about the datasets, even though lines 131 and 284 mention the local scale (650 images) and remote sensing data (600 images).
1) Could you elucidate how the training, validation, and test datasets were delineated using the combined data?
2) Can you specify the categories encompassed within the remote sensing data? For instance, does it encompass iceberg and icebreaker images?
3) When employing different datasets (such as training, validation, and test data), it would be beneficial to display the performance metrics for each set.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report (New Reviewer)
1. The article and the proposed recommendations do not take into account the quality factor of the images obtained, which may contain noise and anomalies, which may lead to a decrease in the reliability of the model provided by the authors.
2. The developed model has not been tested in real conditions, which does not give a full guarantee of its practical value.
Editorial:
1. In the text, the reference to Fig. is used twice. 12 (lines 306 and 310).
2. The order of numbering of tables is broken. Table 3 is followed by Table 5.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 5 Report (Previous Reviewer 4)
see attachment
Comments for author File: Comments.pdf
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report (Previous Reviewer 1)
The last version of the manuscript can be accepted for publication as it is.
Author Response
Thank you very much for taking the time to review this manuscript. It is because of your valuable feedback that the quality of our manuscript can be improved.
Reviewer 3 Report (New Reviewer)
The authors have adequately addressed the comments I made on the manuscript.
Author Response
Thank you very much for taking the time to review this manuscript. It is because of your valuable feedback that the quality of our manuscript can be improved.
Reviewer 5 Report (Previous Reviewer 4)
Accepted
nil
Author Response
Thank you very much for taking the time to review this manuscript. It is because of your valuable feedback that the quality of our manuscript can be improved.
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
The manuscript presents a neural network approach to multi-scale polar object detection. The approach is based on rather well-known algorithms, namely SSD (Single Shot MultiBox Detector) and YOLOv5 (You Only Look Once). It should be noticed that now every day there are appearing hundreds of thousands of articles describing the use of artificial neural networks, "known" or "innovative", which are all written according to the same template and which contain practically no scientific novelty. Unfortunately, the manuscript in question seems to be classified as an article of the type mentioned. As a whole, this manuscript does not seem to have a sound basis, and has many shortcomings that are discussed below.
1. The authors say that the paper has aim to obtain (1) a polar multi-target local scale dataset with 650 images, 5 categories, and 3431 labels, and sea ice, icebreakers, ice melt ponds, inter-ice channels and icebergs identified by the SSD algorithm and (2) a remotely sensed sea ice dataset with a total of 600 images and 15948 labels identified by the YOLOv5 algorithm. Unfortunately, the declared aim implies ordinary routine work and has not led to serious scientific results with scientific novelty that deserve to be published in such a reputable journal as Water.
2. The authors claim that they improve the YOLOv5 algorithm "to perform slicing operations on remote sensing images". However, the reviewer could not find in the text either the justification, or the essence, or a formal description of the changes that improve the algorithm.
3. The "optimization" is also a questionable issue in the paper. What does it mean the phrase: "Combining simulated sea ice images and real sea ice images into a hybrid dataset to verify the accuracy of the optimized YOLOv5 algorithm"? Again, where is justification, essence, and formal description of the "optimized YOLOv5 algorithm"?
4. The article suffers from the main disadvantage inherent in many other articles on neural networks: there is no strict formal justification for the proposed models; the quality of the proposed solutions is justified by a computational experiment, the results of which, given in Section 3.4, are difficult to verify.
5. Many figures many figures are not provided with a sufficient and exhaustive explanation, such as Fig. 6 (what does it mean "masica"); other figures seem to be redundant and does not explain much, for example, Fig. 9.
6. Many sentences are very long. It is recommended to divide them into several short ones. For example,
"In the detection stage, a sliding window is used to rule the specified size such as a 416×416 image as the input of the model, the cut-out adjacent images will have a 15% overlap, and the cropping process is shown in Figure 9, the purpose of this is to ensure that each region is completely detected, although it brings some duplicate detections, the duplicate detections can be filtered out by the NMS algorithm." or
"The largest of these 3, 76´76, is responsible for detecting small targets, which corresponds to the image with resolution 608´608, the perceived field of view of each feature map is only 8´8, and the resolution of remote sensing images is much more than 608´608, some images even reach 16000×16000."
7. The text abounds in stylistic mistakes and typos, therefore, its English checking is highly recommended. For example, the sentence "In this paper, the YOLOv5 algorithm will be improved and the improved algorithm will be applied to the detection of remotely sensed sea ice." is repeated twice; "mosaic" (page 9, line 234) must be "Mosaic", etc.
9. All the reference list must be redeveloped in accordance with the Author's guide.
10. The manuscript, in its present form, cannot be recommended for publication in Water.
See above
Reviewer 2 Report
Review of paper by Shifeng Ding, Dinghan Zeng, Li Zhou, Sen Han, Fang Li, and Jinyan Cai Multi-Scale Polar Object Detection Based on Computer Vision
I find this paper very interesting in the sense of the work to recognize ice on the sea surface. This work has practical applications for navigation in polar regions and will be helpful to avoid ship-ice collisions, such as, for example Titanic in 2012 and many many others. This work will help to detect sea ice and icebergs in advance.
I recommend publication of this manuscript with minor revision. I suggest not using abbreviations in the abstract and conclusions. Some readers read only these sections of publications and will be confused. Thus, the authors should fully write the Single Shot Detector and You Only Look Once algorithms in the Abstract, Key words, and Conclusions.
Eugene Morozov, Shirshov Institute of oceanology, Moscow, Russia
Reviewer 3 Report
In the area of waterways, the author has made a solid effort. However, there is no reading flow from the abstract's initial section to the manuscript's completion. Most of the paper must be rewritten in a more formal yet easy style. The author should put those models to the test for actual route finding so that they can show readers which model is the best to employ
The abstract needs to be rewritten.
Full stops are missing in some places.
Please check the flow of sentences and the tense in which the sentence has to be written
Reviewer 4 Report
see attachment
Comments for author File: Comments.pdf
NIL
Reviewer 5 Report
This is an interesting and timely work. The objectives and methods are well described and verified adequately with the results, thus this study can be effectively used as a reference and guide for polar object detection. The results are presented according to the objectives of the study. However, there are some minor fixes/changes that need to be addressed (as mentioned below).
· In Fig 2 please increase individual figure size and resolution for better visualization.
· In equations 3 and 4, please add the meaning of all symbolic parameters used in equations, for example, r=?, Pinter =? K=? and mAP=?
· In the caption of Figure 5, add what CBL, CSP, SPP and COV stand for.
· In Figure 6, make it a little bigger.
· In Figure 8, increase the image size in the 2nd column
· Recent relevant literature review needs to be added in the Introduction section.
· The discussion part is missing. Critical discussion on the method applied and the finding needs to be added.
· The conclusion only highlights the results, some important limitations and future directions need to be added for the betterment.
