Anomaly-Based Ship Detection Using SP Feature-Space Learning with False-Alarm Control in Sea-Surface SAR Images

Round 1

Reviewer 1 Report (Previous Reviewer 1)

In my opinion, the authors have thoroughly addressed my concerns and I believe that this manuscript is now suitable for publication.

Author Response

Thank you for your professional review.

Reviewer 2 Report (Previous Reviewer 2)

Most of the questions have been responded and the manuscript has been greatly improved. I still have two questions as follows:

 

1. "The size of SP cell can be selected according to the ship size in real SAR images."

This description is still abstract. It is suggested to give actual value of M and example of ship size in the experiments.

 

2. As representatives of deep learning method, YOLO series are not very good at detecting tiny objects. And the hyperparameter setting of the network will affect the results as well. Even so, the comparison results of YOLOv5 are suggested to be supplemented in the manuscript, not only in the responses.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (Previous Reviewer 3)

Dear Authors,

 

You haven’t addressed my previous concerns, so allow me to clarify them here:

1.      You haven’t stated the size of the vessels on which your method is tested as well as environment. Please be specific about the vessel size and environment.

To clarify: Have you used complete HRSID and LS-SSDD-v1.0 datasets or just some portions of those sets? If not, then state which portion of the dataset is used. Are you are targeting in your analyses only large vessels in the blue waters or you considered costal / port / inland waters as well?   

2.      In your reply you have stated that you are using Pfa 10^-5 in your analyses. I must ask you why you do not provide comparative analyses with other methods for that Pfa? Fig 16 shows that for Pfa of 10^-3, but there is no figure which shows behaviour of your method in Pfa 10^{-5 / -6} conditions in comparison to the other methods.

 

Best regards,

Reviewer

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report (New Reviewer)

 This paper focuses on ship detection using Synthetic Aperture Radar (SAR). The authors propose a novel anomaly-based detection method that utilizes feature learning for superpixel processing cells. They extract multiple features, including boundary features, saliency texture features, and intensity attention contrast features, to construct a three-dimensional feature space. The researchers also design an improved clutter-only feature learning strategy to improve target classification accuracy. Experimental results using public datasets demonstrate that the method enhances ship detection performance, achieving high accuracy and low false alarms even in complex environments.        The paper studies the practically important problem of ship detection using SAR images and the experiments are extensive and convincing. I have the following comments to be addressed for the next round of reviews.   1. There are other works that address ship detection using SAR data:   a. Chang, Y.L., Anagaw, A., Chang, L., Wang, Y.C., Hsiao, C.Y. and Lee, W.H., 2019. Ship detection based on YOLOv2 for SAR imagery. Remote Sensing, 11(7), p.786.   b. Rostami, M., Kolouri, S., Eaton, E. and Kim, K., 2019. Deep transfer learning for few-shot SAR image classification. Remote Sensing, 11(11), p.1374.   c. Zhang, T. and Zhang, X., 2019. High-speed ship detection in SAR images based on a grid convolutional neural network. Remote Sensing, 11(10), p.1206.   d. Wei, S., Su, H., Ming, J., Wang, C., Yan, M., Kumar, D., Shi, J. and Zhang, X., 2020. Precise and robust ship detection for high-resolution SAR imagery based on HR-SDNet. Remote Sensing, 12(1), p.167.   e. Liu, N., Cao, Z., Cui, Z., Pi, Y. and Dang, S., 2019. Multi-scale proposal generation for ship detection in SAR images. Remote Sensing, 11(5), p.526.   I think the above works can be discussed in the Introduction section to give a complete perspective to readers about existing works.   2. Could you add more about details of the optimization methods that you use to solve Eq 12? Currently, it is not easy to conclude the approach.   3. In Figure 12 is ti possible to use a different color to highlight the ships that the proposed method is detecting but the prior methods are missing? Currently, it is a bit challenging to distinguish them with visual inspection.   4. Please run your code several times and report both the average performance and the standard deviation on your Table 2 to make the comparison statistically meaningful.      5. Please arrange for releasing the code on a public domain such as GitHub so other researchers can reproduce the results conveniently for future research.   

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report (Previous Reviewer 3)

Dear authors, 

 

You have clarifed the point 1. 

On the other hand, clarification of point 2 (given in the answer to me) is not really convincing.

Since, you have agreed to other reviewer's request to make your code available in a future, I can later check behaviour of your method.  

Please make a note once the code becomes available.

 

Regarding hte paper itself, I require no further actions from your side and believe that the paper may be published in its present form.

 

Best regards,

Reviewer

Reviewer 4 Report (New Reviewer)

The authored have addressed all points raised by the reviewer.

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

Some suggestions are as follows.

(1) Ablation experiments should be used to demonstrate the performance of each component of the proposed method.

(2) It is recommended to add the latest deep learning-based method to comparison methods.

(3) The limitations of your proposed method is not mentioned.

(4) The introduction part should give a briefer INTRODUCTION of the target detection area, rather than introducing the details like L42-106.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The article presents a anomaly-based detection method for ships in SAR images using feature learning for superpixel processing cells. Experimental results show that the proposed method outperforms traditional pixel-level CFAR algorithms. The following concerns are provided to improve the manuscript:

1.       The paragraph beginning at line 42 of the Introduction section, describes the advantages of superpixel segmentation for ship detection and the good performance of the improved method in SAR scenes. However, it seems that the shortcomings of existing methods are not analyzed, making it difficult to highlight the innovation of this paper's method in this section.

2.       The first contribution of the proposed method are based on the boundary feature described by Haar-like, texture feature described by non-uniform LBP and intensity attention contrast feature. However, the Introduction section of the paper does not provide a sufficient review of these features, making it difficult for readers to have a clear background understanding of the development of these three features. From an intuitive perspective, it is also difficult to understand why these three features were selected. It feels like a lack of logical consistency.

3.       The sea-land segmentation and superpixel segmentation is essential for the subsequent steps. How to select the hyperparameters such as S for SP size?

4.       The novelty of the clutter-only feature learning (COFL) model based on anomaly detection decision seems intriguing; however, from a theoretical perspective, its concept is very similar to that of nonlinear SVM, making it difficult to understand its superiority based on the description provided in the Method section. Moreover, the Introduction section lacks references to relevant works that support the so-called "anomaly-based detection decision," giving the impression of a completely new methodology, but the current description fails to convince readers of its effectiveness.

5.       In the experiments, the selection and effects of parameters are not involved. Moreover, the introduction declares that deep learning methods are unfavorable for the fast implementation, however no computation consumption is compared between these two types of methods.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear authors,

 

 

From my point of view your paper is well written and the method presented in it does have some advantages in comparison with the other methods. 

 

In order to further improve the paper, I would suggest the following:

1. Please provide the vessel AIS for the vessels in Figure 11. In this way it will become very clear which type / size of vessels your algorithm is able to detect.

2. Can you provide data regarding the environment in Figure 11, i.e. Sea state and weather conditions. I can understand it might be a little tricky to find that data.

3. Please present behaviour of your algorithm in higher false alarm settings. 10 ^-3 / 10 ^-4 Pfa, presented in the paper (Figure 16)is not very useful in practice, since in majority of maritime applications ask for 10^-5/10^-6 Pfa.

 

Best regards,

Reviewer

The English in the paper is fine.

Apart from some typos, there is not much more to improve.

Author Response

Please see the attachment.

Author Response File: Author Response.docx