MIMO Radar Sparse Recovery Imaging with Wideband Interference Prediction

Round 1

Reviewer 1 Report

In this paper, the authors propose a sparse recovery imaging method with wideband interference (WBI) prediction. The proposed method employs a predictive recurrent neural network (PredRNN) and tensor-based smooth L0 (TSL0) algorithm. From the numerical simulation results presented in the paper, the proposed imaging method has improved performance when compared with others methods in the presence of different WBIs.

 

I consider that the introduction adequately describes the context of the application and presents a summary of the evolution of the existing solutions to solve the WBI prediction problem. However, in my opinion, Sections 2 (Signal Model) and 3 (Conventional and proposed imaging methods) are confusing and should be restructured and rewritten. I list some examples below:

- In Section 2 (Signal Model), many sentences are very long and require more detailed explanations, such as the ones presenting the following equations:

-- Eq. (1), lines 112-119.

-- Eq. (2), lines 120-127.

-- Eqs. (3), (4) and (5) lines 128-138.

-- Eq. (6), lines 139-147

 

- line 138: is it really Figure 2 that should be referenced here? I think it doesn't make sense!

- line 146: matrixes -> matrices

- the entries in Eqs. (7) and (8) are not clear

 

- Section 3 (Conventional and proposed imaging methods) should be much clearer since it presents the main aspects of the operationalization of the proposed method. There are really huge sentences to describe equations which are very dense, for example:

-- Eqs. (11) and (12)

-- All the equations of Sec. 3.2.2.1

Author Response

Thank you very much for your comment.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper proposes a method for dealing with wideband interference. The paper is interesting, but the authors must provide significant changes for publication.

I suggest the authors revise the entire text and try to reduce the passive tense.

The authors must increase the paper length. This journal requires at least 18 pages, and the present paper has 16:

• Articles: Original research manuscripts, which should comprise at least 18 pages with results and discussion separately in two sections. The journal considers all original research manuscripts provided that the work reports scientifically sound experiments and provides a substantial amount of new information. Authors should not unnecessarily divide their work into several related manuscripts, although short Communications of preliminary, but significant, results will be considered. The quality and impact of the study will be considered during peer review.

I suggest the authors write the paper in a more conventional flow. Please use the section names suggested by this journal: Introduction, Materials and Methods, Results, Discussion, and
Conclusion.

The paper has no discussion. This is a very important topic that the authors can explore to really show the scientific community the advances of the proposed methodology. This section can also help to reach the minimum number of pages.

All figures and tables must be self-explanatory. Thus, all abbreviations and acronyms must be described in the caption.

Some figures are very hard to see, especially the axis names. I suggest making larger labels or larger figures.

In most deep learning studies, it is conventional to provide certain information that is currently missing in this paper, such as parameter choice for the deep learning models, the dataset specifications, and how the data was split in training, validation, and testing:

·         There is no explanation regarding the parameters (learning rate, batch size, etc.)

·         There is no explanation on the dataset (e.g., number of samples)

 

·         There is no explanation of how the testing was performed.

Author Response

Thank you very much for your comment. 

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

A novel method for MIMO radar recovery imaging is presented in this paper. The main part of proposed method is the prediction of Wideband Interference. Author’s proposal is to use Predictive Recurrent neural network for this task, and throughout the manuscript they present mathematical background and give test results and robustness and accuracy analysis. The topic is very interesting and, as authors have shown, produces promising results. The idea of using predicted instead of historical WBI in processing of radar signal is very good, as well as the usage of PRNN for prediction. The paper is well structured, scientifically solid and easy to read. References are appropriate, relevant and well cited.

I have not found any major drawbacks of this paper. Manuscript is free of typing errors, acronyms are explained and grammar is good. The only remark would be to the size of figures 9 and 10. These figures show the results of applied methods and they are too small. Although most important figures are figures 11 and 12, still figures 9 and 10 are important too. They are well organized already, but being so small they lose their purpose. Maybe authors should reduce the number of sub-figures, and show the most interesting ones followed by appropriate comments.

Author Response

Thank you very much for your comment. 

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The paper is well presented from the radar point of view, contrary to the part of the Predictive Recurrent Neural Network (PredRNN). Once completed, the paper will be interesting.

I propose to reconsider the publication of the paper after enhancing the part of the Neural Network: optimizer, learning rate strategy, maximum number of epochs, database size for training and validation, curves of training and validation accuracy/losses, use of a specific library (Pytorch, Keras, proprietary?).

Additional minor points to be considered: 

- Have you tried an interference to signal ratio (ISR) of less than 40dB, for example 10dB or less?

- Correct word "hencing" in page 9, line 295

- Add vertical line at middle of Table 1

Author Response

Thank you very much for your comment. 

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

In this paper, the authors propose a sparse recovery imaging method with wideband interference (WBI) prediction. The proposed method employs a predictive recurrent neural network (PredRNN) and tensor-based smooth L0 (TSL0) algorithm. The authors present numerical simulation results to argue that the proposed imaging method has improved performance when compared with others methods in the presence of different WBIs.

In the previous review, I made some suggestions and comments about general aspects of the manuscript. Many changes were adequately carried out by the authors in the current version of the paper. However, some additional adjustments can be made to provide a better understanding of the operationalization of the proposed method.

Next, I list some observations on remaining issues for the authors to consider before submitting the final version of the manuscript:

- The sentence in lines 81-84 is not comprehensible.

- The sentence in lines 146-154 presents important definitions to the understanding of the "signal cube". However, it is very long and confusing.

- The sentence in lines 155-159 is very long and confusing. Additionally, note that  $\times_1$, $\times_2$, and $\times_3$ denote mode-tensor by matrix products and the corresponding statement about notation should come after Eq. (9) and not after Eq. (11).

- In my opinion, the structure of Sec. 2.1 should be improved. Note that, in this section, not only the signal model is introduced, but also the problem formulation is presented, as can be observed from Eqs. (11), (12), and (13) (i.e. optimization problem to be solved).

- Add the meaning of acronyms SRI (sparse recovery imaging) and LSTM (Long Short-Term Memory)

- The equations in Sec. 2.3.2 are very dense, especially (19) and (20). Are all the variables scalars in these expressions? In my opinion, it is very important to employ adequate notation to facilitate the understanding of the steps of the proposed method.

- In Eq. (23), it seems parentheses are missing because the operation "/D_X/Z" can lead to ambiguity. Do the authors mean "argmin(../D_X)/Z" ?

- In general, the paper presents long and confusing sentences discouraging the reader to continue reading. This, as a consequence, demeans the work itself. I recommend that the authors do a very careful reading and reviewing of the text, avoiding very long sentences and trying to give all clues to facilitate the understanding of the proposed method operation and implementation. 

Reviewer 2 Report

The authors attended to all of my requests.

Reviewer 4 Report

Thank you for the new improved version. Although I am still puzzled by the performance of the NN. The authors only show the training loss. Important parameters are also the validation loss and accuracies for the training and the validation.

I propose to accept the paper for publishing after these minor points are included.