Design of Siamese Network for Underwater Target Recognition with Small Sample Size
Round 1
Reviewer 1 Report
The manuscript “Design of Siamese Network for Underwater Target Recognition with Small Sample Size” tries to bridge the gap between the requirement of large training samples for acceptable accuracy and the increased associated cost particularly, in underwater target recognition. The authors claimed to design datasets with different parameters and employed a Siamese network for this task. Though many works have already been reported in this domain yet, the findings look promising. My observations regarding the manuscript:
1. The abstract should briefly incorporate the notable improvements relative to the reported work.
2. The authors should highlight the need/importance of the proposed work and clearly mention the objectives and novelty of the work based on the reported literature and research gap.
3. The authors are suggested to append more literature that contains the existing techniques for the task along with their bottlenecks.
4. How LPF restricts sin(Ω-2ω)t and cos(2Ω-2ω)t in Eq. (4)? Justify.
5. Why the sampling rate of the ship-radiated noise was set to 44.1 kHz in section 2.2.1? Justify.
6. What does C23000 represent in line no 155 page no 5.
7. Eq. 9 and 10 need to be rechecked.
8. The authors should justify the claim “Network performance no longer improved significantly with more than three convolution layers” in line no 201, page 7.
9. Why did the authors stick with ReLU activation function when there are a number of activation functions found in the literature? Also, how did the authors manage the information loss due to the employment of ReLU?
10. Did the authors use any warm-up epochs?
11. The authors should present a comparative analysis to validate the performance of the proposed work.
12. The authors are also encouraged to mention the difficulties associated with the approach and how they mitigate it.
13. The conclusion should be followed by directions for future work.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors have developed the Siamese network for identifying DEMON spectra of underwater targets with a limited sample size. Some of the observations are as follows:
1. Some results can be added in the abstract to highlight the scope of the manuscript.
2. Currently literature part is not strong. I suggest you add the literature with the most recent references in the introduction section.
3. Page 2, “Various DEMON spectra were used 56 to construct the datasets to train the network.” Such as ….
4. Page 2, “Due to the cost of underwater target data acquisition, this study developed, trained, and evaluated a Siamese network using simulated data.” Revised this line… simulation and actual acquisition are different and both are necessary.
5. Revise this line as “ship-radiated noise L(t) can be written as”
6. In Figure 1, input and output are missing.
7. On Page 4, Figure2, the x-axis unit is missing. Correct it “frequency (in Hz)”
8. Page6, why was this CNN selected as compared to others? Add a few lines.
9. Figures 5-10 need to be enlarged.
10. Limitations or challenges faced need to be specified in the discussion.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
