Automatic Radar Intra-Pulse Signal Modulation Classification Using the Supervised Contrastive Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the authors propose a method leveraging contrastive learning for intra-pulse modulation recognition of radar signals. This attempt aims to achieve stronger modulation recognition capabilities with only a limited number of labeled samples within a large dataset. It has the potential to facilitate subsequent radar waveform feature extraction and interference strategy formulation. However, in my opinion, there are still several aspects of this paper that require revision for publication.

1.  On line 22, the authors mention that with the advancement of radar technology, radar signal detection has become more difficult. Moreover, waveform modulation recognition is typically performed after radar signals have been detected. Consequently, does the degraded detection performance of radar pulses impact modulation recognition? Like less samples or poor sample quality.

2.  On lines 27 and 28, the authors claim that modulation recognition or classification benefits radar detection, which may inaccurate. As I mentioned in point 1, classification or recognition are generally conducted after signal detection. Without first detecting the signal, classification becomes impossible. Additionally, the term "radar detection" is vague; clarification is needed to specify whether it refers to "the target detection of radar" or "the detection of radar signal."

3.  Regarding the proposed method, its novelty appears limited. According to the authors' description, they merely utilize the time-frequency analysis results (CWD images) of signals, as an input into a two-stage combined network. It remains unclear what specific efforts the authors made to enhance performance under limited labeling conditions and what problems this approach can address. The innovation points and contributions of this paper are not clearly articulated.

4.  Based on the abstract and introduction, the core objective of this paper is to achieve stronger waveform modulation classification or recognition under conditions of limited labeled samples. Consequently, the authors should discuss the difference in classification accuracy between scenarios with 1% and 10% labeled samples, with a fixed sample size. An experimental analysis of label proportion versus prediction accuracy is essential.

5.  Are the parameters in Tables 7 to 10 fixed, or do they need to be adjusted for different sets of samples? If these optimal parameters vary with the sample, considering that reconnaissance-acquired radar signal samples are often numerous and unknown in practical applications, does the proposed algorithm require adaptive optimization for each collected sample set?

6.  In Figure 3, are the NLFM waveform in (b) and the Frank waveform in (d) mislabeled?

 

Comments on the Quality of English Language

none

Author Response

Please see attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Suggested Enhancements to the Presentation of Experimental Results:

1. FIG3 Adjustments: Add horizontal and vertical coordinate labels to FIG3. Additionally, verify the accuracy of the time-frequency figure of LFM.
2. Graphical Representation: Replace the extensive use of tables with graphs to improve clarity and visualization.
3. Result Details: Include more detailed descriptions of the results, such as illustrating the change in accuracy with the number of epochs.

Comments on the Quality of English Language

 Minor editing of English language required.

Author Response

Please see attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

 

 

For radar intra-pulse signal modulation classification, a supervised learning model using the contrastive learning method (SL-CL) is proposed in this manuscript. The experiments show that the SL-CL model outperforms the comparison models in the situation of a small number of labeled samples. Some issues should be improved as follows:

 

1.      The literature review is somewhat limited and could be expanded to include recent work on radar/communication signal modulation classification.

2.      The main contribution and novelty of this paper can be further highlighted in the introduction.

3.      As the paper said “SeSL has high time consuming for processing the pseudo labels of the unlabeled samples”, but the paper did not compare the time complexity between SL-CL with SeSL.

4.      In line 209, it said “where U(.) represents a uniform random distribution”, but in Table 5, we can not see this function.

5.      The paper said the proposed SL-CL method can achieve a high accuracy with limited labeled samples, and the experiments under 25,40 and 50 training samples per class are carried out. So, it would be nice to have performance bounds on the number of samples.  That is, the sample size requirement needs to be further clarified.

6.      The reason behind the low classification accuracy of signals P1 and P4 should be analyzed.

7.      It would be better that the comparison methods in the experiments including recent works, as the SIMCLR was proposed in 2020.

 

 

Overall, I recommend major revisions for this paper. 

Author Response

Please see attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

The author addresses the challenges of capturing high-quality radar signals and the limited number of labeled samples in the field of radar intrapulse signal modulation recognition by proposing the SL-CL model. This model employs a two-stage training process using supervised contrastive loss and cross-entropy loss, enhancing feature extraction capabilities and offering a viable approach to radar intrapulse signal modulation recognition. However, certain aspects of the study require improvement:

1.       The abstract outlines the research background and issues, covering the main aspects of the study, but should further elaborate on the research methods and the results achieved.

2.       The introduction lacks an in-depth exploration of existing methods and should be more systematic. It should provide additional information on the applications of contrastive learning in radar intrapulse signal modulation classification techniques, clearly identifying the limitations and shortcomings of current research. This will highlight the innovation of this study. The main contributions and innovations of this paper should also be described in detail.

3.       The architecture of the SL-CL model shown in Figure 1 is inaccurate. The model is trained in two stages. In the second stage, the encoder parameters are fixed and the projection network is discarded, allowing the signals to flow directly into the classification network after encoding. This is not consistent with what is visually depicted in the figure.

4.       The author does not provide a detailed network parameter structure for the SL-CL model. Please include a comprehensive description of the network parameters and training strategies.

5.       VGG is a classic CNN feature extraction network in the field of computer vision. However, the author has labeled VGG as a two-stage training model in Table 6. Please explain the rationale behind this classification in detail.

6.       The network models used in the model comparison experiments in Table 11 mainly consist of classic transfer learning models. Please include additional, more convincing network models to highlight the advantages of the proposed network.

Author Response

Please see attachment.

Author Response File: Author Response.docx

Reviewer 5 Report

Comments and Suggestions for Authors

A light deep learning model for the radar signal modulation classification based on CL and SL is proposed in this   manuscript. According to the results, the model has certain effects, but from the experimental process, the model's ability to process higher-resolution images is limited. Can the author supplement relevant comparative experiments based on the existing model?

 

And, attention should be paid to the analysis of experimental results according to the characteristics of model construction. From the perspective of this manuscript, only the model construction and experimental execution were completed, and more detailed analysis of experimental results was lacking in order to series the whole manuscript

 

In addition, it is suggested that the Manuscript be changed to paper in the revised version, which would be more appropriate in a scientific research paper

 

If the content font size in Table 8 and 9 is incorrect and smaller than the body size, the author should determine whether the problem exists in other tables.

 

 

The author needs to discuss the experimental results in more detail, and should not only introduce the improvement of the effect, but also discuss the advantages and disadvantages of SL-CL model in the conclusion.

Author Response

Please see attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

none

Comments on the Quality of English Language

none

Reviewer 3 Report

Comments and Suggestions for Authors

All issues raised by the reviewer have been revised.

Reviewer 5 Report

Comments and Suggestions for Authors

All the questions have been answered and met the standards for publication.