Small Floating Target Detection Method Based on Chaotic Long Short-Term Memory Network

Round 1

Reviewer 1 Report

In this paper, a methodology for weak signal detection using LSTM is proposed. 
In this context, the Complementary Ensemble Empirical Mode Decomposition (CEEMD) method is used to denoise the signal.
After studying the manuscript and the related references the following comments are stated:

1) The language needs improvement.
2) The contribution and novelty of this work over the existing methods are not defined clearly.
3) The main weakness of this work is the absence of any comparison with similar methods from the literature.

Author Response

First of all, thank you very much for your comments on the content of the article. We believe that your comments will be of great help to the academic level of the article.

We have carefully considered your comments and made corresponding modifications in the article. The following is the response to your comments

Point 1:The language needs improvement.

Response 1: We proofread the article carefully and revised many language mistakes, such as standardizing the abbreviation format and correcting the wrong words in the article.

Point 2: The contribution and novelty of this work over the existing methods are not defined clearly.

Response 2: In this paper, the deep learning LSTM network is applied to the field of small target detection in the background of sea clutter for the first time, which is a data-driven detection model.The purpose of this paper is to utilize the powerful learning ability of deep learning to develop new methods for radar detection under the background of big data.It solves the problems of over-dependence on mathematical model and poor real-time performance of existing methods.

We have added the following to the Abstract:The proposed chaotic long and short term memory network, which determines the training step length according to the width of embedded window, is a new detection method, which can accurately detect small targets submerged in the background of sea clutter.

Point 3: The main weakness of this work is the absence of any comparison with similar methods from the literature.

Response 3: Thank you for pointing out this problem. After careful consideration, we would like to explain to you:

The article has done two main innovation work.The first point is that the CEEMD-WPT denoising algorithm conforming to the sea clutter property is proposed. This method is an improvement on the CEEMD algorithm. In order to compare the denoising effect, a 10 dB white noise is added to the sea clutter data, and the complementary integrated empirical mode decomposition (CEEMD) and empirical mode decomposition (EMD) denoising algorithms are compared respectively, The effectiveness of the proposed denoising algorithm is verified. The experimental results are shown in Table 1：

Table 1. Comparison of denoising signal to noise ratio.

The second innovation is to introduce the deep learning LSTM network into the field of sea clutter detection, modify the network parameters according to the chaotic characteristics of sea clutter, and finally complete the task of small target detection. the network designed in this paper is driven by original data and does not require dimension transformation and domain change in advance, which is different from the detection methods based on fractal statistics and FRFT domain listed in the literature.According to the data characteristics, the model is optimized by iterative learning through gradient descent algorithm and directional propagation algorithm.

The methods in the literature are different from the design idea of this paper, and the detection indexes are also different. The design method focuses on detecting small targets from sea clutter and finding undetected sub-targets by improving network parameters.

You can also see the response letter more clearly in the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This article is about the small floating target detection method using chaotic long short term memory network. I have some detailed comments to help improving this article. 

1, 'long short term memory' should be addressed clearly in both abstract and the main text to avoid any confusion.

2, The abstract should highlight the unique features of this method. The accuracy is improved by 44%, what is compared with this new method? 

3, Figure 1 has so many information, please consider to divid this figure into two figures. 

4, Table 1 keep 4 decimals, Table 2 keeps 6 decimals, please consider to use 4 decimals or even 3 through out the article.

5, What are the more application of this method? It should be elaborated regarding what are the next step to improve this method? What are the difference between this method/results and other methods/results in the literature? 

Author Response

First of all, thank you very much for your comments on the content of the article. We believe that your comments will be of great help to the academic level of the article.

We have carefully considered your comments and made corresponding modifications in the article. The following is the response to your comments.

Point 1: 'long short term memory' should be addressed clearly in both abstract and the main text to avoid any confusion.

Response 1: We have proofread the full text and confirmed that all 'Long Short Term Memory' are clearly explained.

Point 2: The abstract should highlight the unique features of this method. The accuracy is improved by 44%, what is compared with this new method? 

Response 2:  The article has done two main innovation work.The first point is that the CEEMD-WPT denoising algorithm conforming to the sea clutter property is proposed. It can be seen from Table 1 that the SNR of the proposed new method is improved by 33.6% on average

Table 1. Comparison of denoising signal to noise ratio.

The second innovation is to introduce the deep learning LSTM network into the field of sea clutter detection, modify the network parameters according to the chaotic characteristics of sea clutter, and finally complete the task of small target detection.

We have added the following to the Abstract:The experimental results show that the CEEMD-WPT algorithm is consistent with the target distribution characteristics of sea clutter, and the denoising performance is improved by 33.6% on average. The proposed chaotic long and short term memory network, which determines the training step length according to the width of embedded window, is a new detection method, which can accurately detect small targets submerged in the background of sea clutter.

 Point 3: Figure 1 has so many information, please consider to divid this figure into two figures. 

Response 3: Thank you for pointing out this problem. We have divided Figure 1 into two graphs to more clearly illustrate the proposed algorithm.Please see Figures 1 and 2 in the article

Figure 1. CEEMD-WPT denoising algorithm flow chart.

Figure 2. LSTM detection method of weak signal in chaotic sequence.

Point 4: Table 1 keep 4 decimals, Table 2 keeps 6 decimals, please consider to use 4 decimals or even 3 through out the article.

Response 4: We have unified the format of the numbers in the table, keeping three decimal  and correcting other language errors.

Point 5: What are the more application of this method? It should be elaborated regarding what are the next step to improve this method? What are the difference between this method/results and other methods/results in the literature? 

Response 5: The proposed approach can be used for a long period of time sequence forecast, such as data mining, the urban road network traffic flow prediction, photovoltaic power station irradiation intensity prediction and other fields.The next step we are going to use the latest measured radar data, design optimization algorithm to improve other LSTM network parameters, establish evaluation index based on deep learning detection mechanism.

As for your question about the difference between this paper and other methods in the literature, we would like to make the following explanation:In this paper, the deep learning LSTM network is applied to the field of small target detection in the background of sea clutter for the first time, which is a data-driven detection model.The purpose of this paper is to utilize the powerful learning ability of deep learning to develop new methods for radar detection under the background of big data.It solves the problems of over-dependence on mathematical model and poor real-time performance of existing methods.

We added the following to the conclusion:The method proposed in this paper is not only effective for small target detection in the background of sea clutter, but also can be used for the prediction of long time series in other fields. The next step we are going to use the latest measured radar data, design optimization algorithm to improve other LSTM network parameters, establish evaluation index based on deep learning detection mechanism.

You can find the response letter clearly in the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The presentation of the paper needs to be improved. There are some concerns regarding the overtraining of the model. Authors need to address how their model is not overfitting.

 Line 35:

 CEEMD already defined.

Acronyms not defined:

Line 49: Define “BP”

Line 58: LSTM

Line 71: C-C

98: EMD

Line 246: SNR

Line 66–68: Why did you consider low-frequency as the main and the high as noise? What are the merits? Has any other study done this? If not, support your argument.

Line 91: Explain the significance of the white noise.

Line 129–131: Sentence not clear.

Line 148: How do you make sure that the model is not overtrained? Explain how you would prevent overfitting.

Figure 1: “Testing Ste”, typo?

Line 216: “cemd”, typo?

Table 1: Not sure what the Chinese text is.

Figure 6: The individual figure captions (Line 268) is not very clear.

Line 295–297: Not very clear. What kind of contingency? Explain.

Table 2: Remove one of the table captions.

Line 312–318: How is this different from model overtraining? When the model is overtrained, it works better but cannot get a generalized result? Please address this.

Author Response

Thank you for your valuable advice.We have revised the content of the article according to the advice raised by the experts.

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper has presented interesting findings using Wavelet Packets Denoising and Chaotic Long Short-Term Memory Network to detect small floating targets. The research has used Wavelet packets for denoising the signal. Have the author's considered other wavelet techniques such as Discrete wavelet transform, Stationary Wavelet Transform or Dual-Tree Complex Wavelet Transform for denoising. It would be interesting to see how your algorithm works with abrupt changes in sea conditions to detect small objects.

The paper just very minor amendments before publication:

1. Use of the word be consistent using either denoising or de-noising do not mix the word.
2. There are many places in the paper a space needs to be added, i.e. before the reference back or after it and after the period (or full stop).
3. Also, you use many abbreviations throughout the paper without stating it first, i.e. LSTM is Long Short-Term Memory, so where you first use it, say Long Short-Term Memory (LSTM). A technical reader/expert would understand it, but it helps to reinforce it for a novice reader.

Author Response

First of all, thank you very much for your comments on the content of the article. We believe that your comments will be of great help to the academic level of the article.

We have carefully considered your comments and made corresponding modifications in the article. The following is the response to your comments.

Point 1:Use of the word be consistent using either denoising or de-noising do not mix the word.

.Response 1: We have proofread the full text and revised the following parts：

• “The denoised high frequency IMFs and low frequency IMFs reconstruct the pure sea clutter signal together.”
• “The high frequency part is denoised by wavelet packet transform(WPT).”

Point 2: There are many places in the paper a space needs to be added, i.e. before the reference back or after it and after the period (or full stop).

Response 2: We have reviewed the full text and corrected the format errors to make the article more readable.

Point 3: Also, you use many abbreviations throughout the paper without stating it first, i.e. LSTM is Long Short-Term Memory, so where you first use it, say Long Short-Term Memory (LSTM). A technical reader/expert would understand it, but it helps to reinforce it for a novice reader.

Response 3: We are very sorry for the bad reading experience brought to you due to our mistake.We have proofread the full text and revised the following contents:

1)“Among them, Back Propagation (BP) neural network is widely used as a typical network in artificial intelligence methods”

2)“Wu et al.[15] proposed a new method for fault prediction of equipment degradation sequence based on Long Short-Term Memory(LSTM) network by taking advantage of LSTM's long-term dependent learning ability.”

3)“ Each signal in the set is decomposed by empirical mode decomposition (EMD)”

4)“According to table 1, the signal-noise ratio(SNR) of CEEMD-WPT de-noising algorithm based on the new method of intrinsic mode function extraction is the highest.”

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed all the comments of the reviewer. The revised manuscript has been improved.

Reviewer 2 Report

this version is fine.