Trunk Borer Identification Based on Convolutional Neural Networks
Round 1
Reviewer 1 Report
This research proposes a CCN named TrunkNet to detect the activity of an insect called Agrilus planipennis Fairmaire in the trunk by recording and analyzing its vibration signals in audio form using a developed sensor by the researchers. The CCN offers higher performance compared with the previous works in the research area. My recommendation for further improvements are mentioned below:
1- In the Abstract, please add the limitation(s) of the proposed work and provide a concise suggestion for future work.
2- In the Introduction, please add the last paragraph to explain briefly the following sections of the paper.
3- The manuscript lacks the inclusion of relevant and related literature. Then, the Related Works section should be added after the introduction in which a review of the recent studies related to this work should be conducted, and a comparison between the present study and them should be performed (Pros and cons perspectives). This review could help the readers to point out the current research gap(s) and define their research objectives, accordingly.
4- Please mention that the cylinders are the trunk of the tree, In Figure 2.
5- Please improve the title of Table 1.
6- The quality of Figure 4 should be improved.
7- Please check the parameter p in equation 1 and the text. There is inconsistency.
8- The manuscript requires adequate referencing for example in 3.2.
9- The authors are encouraged to state the entire phrase before using any abbreviations. Adding an abbreviation section to the end of the paper before the reference section is also beneficial for readers to refer to.
10- Please add the link in the paper to the source code and the dataset to increase the reproduction of the study by other scholars.
11- Please add other parameters in the results to examine the performance of each neural network, such as execution time.
12- Please attempt to use references to established journals and conferences rather than preprint depositories that publish non-reviewed research.
13- The current submission requires noteworthy proofreading, as several incomplete, and unclear sentences hinder demonstrating the importance of the work. I highlighted some concerns about the English language in the file attached.
14- For other corrections, please refer to the attached file.
15- Please highlight the corrections for my comments in the revised file to speed up the review process.
Best wishes.
Comments for author File: 
Comments.pdf
Author Response
Please review the submitted Word.
Point 1: In the Abstract, please add the limitation(s) of the proposed work and provide a concise suggestion for future work.
Response 1: We refined the abstract and put forward suggestions for future research.The abstract is shown below.
Abstract: Trunk borer is a great danger to forests because of its strong concealment, long lag and great de-structiveness. In order to improve the early monitoring ability of trunk borer, the representative Agrilus planipennis Fairmaire was selected as the research object. The convolutional neural net-work named TrunkNet was designed to identify the acitvity sounds of Agrilus planipennis Fair-maire larvae. The activity sounds were recorded as vibration signals in audio form. The detector were used to collect the activity sounds of Agrilus planipennis Fairmaire larvae in the wood seg-ments and some typical noise outdoor. The vibration signal pulse duration is short, random, and high energy. TrunkNet was designed to train and identify vibration signals of Agrilus planipennis Fairmaire. Over the course of the experiment, the test accuracy of TrunkNet was 96.89%, while MobileNet_V2,ResNet18 and VGGish showed 84.27%, 79.37% and 70.85% accuracy, respectively. TrunkNet based on the convolutional neural network can provide technical support for the au-tomatic monitoring and early warning of the stealthy tree trunk borers. The work of this study is limited to a single pest. The experiment will further focus on the applicability of the network to other pests in the future.
Point 2: In the Introduction, please add the last paragraph to explain briefly the following sections of the paper.
Response 2: We have added a part in the introduction to introduce the composition of the paper and the experimental process.
The rest of this study is organized as follows: Section 2 introduces the current related research work. Section 3 describes the method of data collection, data processing and data set division. In section 4, a network model for identifying activity vibration signals of Agrilus planipennis Fairmaire is designed. Section 5 describes and analyzes the experimental results. Section 6 summarizes this paper.
Point 3: The manuscript lacks the inclusion of relevant and related literature. Then, the Related Works section should be added after the introduction in which a review of the recent studies related to this work should be conducted, and a comparison between the present study and them should be performed (Pros and cons perspectives). This review could help the readers to point out the current research gap(s) and define their research objectives, accordingly.
Response 3: We have added the "Related Work" section. In the relevant fields of this study, we summarized the development and progress for readers to think.
Point 4: Please mention that the cylinders are the trunk of the tree, In Figure 2.
Response 4: We have revised the paragraph to increase readers' understanding.
Point 5: Please improve the title of Table 1.
Response 5: We have revised the title of Table 1. The new title is “Signal details”. It is shown below:
Table 1. Signal details
|
Categories |
Duration(h) |
Sample rate(kHz) |
Sample depth(bit) |
|
Agrilus planipennis Fairmaire vibration signals |
32 |
44.1 |
16 |
|
Non-vibration signals |
10 |
44.1 |
16 |
Point 6: The quality of Figure 4 should be improved.
Response 6: We have revised the Figure 4. It is now clearly visible and shown below:
Point 7: Please check the parameter p in equation 1 and the text. There is inconsistency.
Response 7: We made changes to ensure consistency. It is shown below:
|
(1) |
where the σ(x) is exponential linear cell activation function, x is the input value, and the ρ hyperparameter can control when the negative part of the output of the activation function is saturated, which is taken as 1.0 in this study.
Point 8: The manuscript requires adequate referencing for example in 3.2.
Response 8: We have made corrections in the “Identification Model” paragraph, indicating the relevant literature for inspiration.
Point 9: The authors are encouraged to state the entire phrase before using any abbreviations. Adding an abbreviation section to the end of the paper before the reference section is also beneficial for readers to refer to.
Response 9: For the readers' reading experience, we have replaced the abbreviation of the article as the full name to prevent unnecessary misunderstanding.
Point 10: Please add the link in the paper to the source code and the dataset to increase the reproduction of the study by other scholars.
Response 10: We added a link in the appendix of the article.
Point 11: Please add other parameters in the results to examine the performance of each neural network, such as execution time.
Response 11: We added F1-score as parameter for network performance comparison. It is shown below:
Table 3. Recognition Result
|
Recognition Nets |
Accuracy Rate(%) |
F1-score |
|
TrunkNet |
96.89 |
0.92 |
|
MobileNet_V2 |
84.27 |
0.86 |
|
ResNet18 |
79.37 |
0.69 |
|
VGGish |
70.85 |
0.58 |
Point 12: Please attempt to use references to established journals and conferences rather than preprint depositories that publish non-reviewed research.
Response 12: We have corrected and updated the references.
Point 13: The current submission requires noteworthy proofreading, as several incomplete, and unclear sentences hinder demonstrating the importance of the work. I highlighted some concerns about the English language in the file attached.
Response 13: According to your suggestions, we rewrote and revised some sentences to ensure the reading quality.
Point 14: For other corrections, please refer to the attached file.
Response 14: We carefully reviewed the attached file and made corrections one by one.
Point 15: Please highlight the corrections for my comments in the revised file to speed up the review process.
Response 15: In the newly uploaded manuscript, we have completed the revision where you put forward your suggestions. Please refer to the annotation mode for details.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This manuscript proposed a novel deep learning-based approach for trunk borer identification, where a convolutional neural network (CNN) was developed for the task of interest. First of all, the audio signals were collected by piezoelectric transducer, which were then trained and tested by the deep learning network. Finally, the performance of the proposed method was validated using experimental data, with satisfactory results. Overall, the topic of this study is interesting, and the manuscript was well organised and written. It can be considered to be accepted for publication in Applied Sciences, if the authors can well address the following comments.
1. The main innovation and contributions of this study should be clearly clarified in both abstract and introduction.
2. Broaden and update literature review on CNN or deep learning in practical engineering applications. E.g. Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network; Vision-based concrete crack detection using a hybrid framework considering noise effect.
3. How did the authors set the hyperparameters of the proposed network to achieve the best prediction result? In general, the deep learning model’s performance is mainly related to hyperparameter setting.
4. This study only used accuracy as the evaluation metric. Please consider more to give a comprehensive evaluation and comparison.
5. Several pretrained CNN models, such as ResNet, MobileNet, etc, were used to be compared with proposed one. However, these deep CNN models were originally developed for other task. Accordingly, transfer learning is required. Please add more information about how these models were transferred to the ones used for the task in this research.
6. How about the robustness of the proposed method against noise effect?
7. More future research should be included in conclusion part.
Author Response
Please review the submitted Word.
Point 1: The main innovation and contributions of this study should be clearly clarified in both abstract and introduction.
Response 1: We revised the abstract and introduction, and described the main contributions of the study in relevant positions. The abstract is shown below.
Abstract: Trunk borer is a great danger to forests because of its strong concealment, long lag and great de-structiveness. In order to improve the early monitoring ability of trunk borer, the representative Agrilus planipennis Fairmaire was selected as the research object. The convolutional neural net-work named TrunkNet was designed to identify the acitvity sounds of Agrilus planipennis Fair-maire larvae. The activity sounds were recorded as vibration signals in audio form. The detector were used to collect the activity sounds of Agrilus planipennis Fairmaire larvae in the wood seg-ments and some typical noise outdoor. The vibration signal pulse duration is short, random, and high energy. TrunkNet was designed to train and identify vibration signals of Agrilus planipennis Fairmaire. Over the course of the experiment, the test accuracy of TrunkNet was 96.89%, while MobileNet_V2,ResNet18 and VGGish showed 84.27%, 79.37% and 70.85% accuracy, respectively. TrunkNet based on the convolutional neural network can provide technical support for the au-tomatic monitoring and early warning of the stealthy tree trunk borers. The work of this study is limited to a single pest. The experiment will further focus on the applicability of the network to other pests in the future.
Point 2: Broaden and update literature review on CNN or deep learning in practical engineering applications. E.g. Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network; Vision-based concrete crack detection using a hybrid framework considering noise effect.
Response 2: According to your suggestion, we corrected the relevant literature review.
Point 3: How did the authors set the hyperparameters of the proposed network to achieve the best prediction result? In general, the deep learning model’s performance is mainly related to hyperparameter setting.
Response 3: In the corresponding paragraph, we give the adjustment method of the super parameter for the reader's reference. It is shown below:
Methods for setting hyperparameters include manual tuning and automatic tuning. Based on the related experiments of audio recognition and the workload of this research, we adjusted the hyperparameters manually. After a lot of experimentation, we found the best performing parameter size and used it.The initial learning rate was set to 0.0003 in the whole process, using the cross-entropy function as the loss function. Meanwhile, batch size is 64. The model training ended after 1100 rounds.
Point 4: This study only used accuracy as the evaluation metric. Please consider more to give a comprehensive evaluation and comparison.
Response 4: We added F1-score as a parameter for network performance comparison. It is shown below:
Table 3. Recognition Result
|
Recognition Nets |
Accuracy Rate(%) |
F1-score |
|
TrunkNet |
96.89 |
0.92 |
|
MobileNet_V2 |
84.27 |
0.86 |
|
ResNet18 |
79.37 |
0.69 |
|
VGGish |
70.85 |
0.58 |
Point 5: Several pretrained CNN models, such as ResNet, MobileNet, etc, were used to be compared with proposed one. However, these deep CNN models were originally developed for other task. Accordingly, transfer learning is required. Please add more information about how these models were transferred to the ones used for the task in this research.
Response 5: We explained how to apply these networks to this experiment and how we operate in the text. It is shown below:
To verify the recognition accuracy of TruckNet, we use some mature network struc-tures for comparative experiments. In convolutional neural networks, as the number of layers increases, so does its ability to fit more complex functions. Neural networks use a backpropagation algorithm for weight updates, but increasing the number of layers will eventually cause the gradient to vanish. The short-circuit connection of ResNet is very useful for how to avoid the problem of gradient disappearance and explosion. The depth-wise separable convolution of MobileNet_V2 expands the feature map channel and en-riches the number of features by introducing residual blocks. This approach significantly improves the recognition accuracy. In this study, in order to test the proposed network ef-fect, we introduced ResNet, MobileNet and other network comparisons through transfer learning. We use the same dataset and data processing. We modify the last layer of the network model and set its last layer as a new classification layer. In this way, the network can ensure that the output dimension is the same as the number of types in the data set, and then the network is evaluated. So we use VGGish, ResNet and MobileNet as the com-paired network to recognize Agrilus planipennis Fairmaire vibration signals.
Point 6: How about the robustness of the proposed method against noise effect?
Response 6: In the data set design stage, in order to greatly simulate the real natural environment, we have conducted additive mixing of noise and signal as the signal that really participates in the experiment. Then we identify the mixed signal and noise. In this case, the environmental noise does not affect the performance of the network.
Point 7: More future research should be included in conclusion part.
Response 7: We modified the conclusion and added the idea of future research. It is shown below:
In the future, we will collect vibration signals from different forest areas to further verify the feasibility of convolutional neural networks. We will pay more attention to the population density information of pests to determine the specific damage of trees.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Adequate modifications have been considered. English language and style are fine, and minor proofreading is required.
Reviewer 2 Report
All the comments have been addressed.
