Application of Machine Learning for Insect Monitoring in Grain Facilities

Round 1

Reviewer 1 Report

The manuscript describes technological development concerning machine learning of stored product insects. The methods concerning the technology are well described. There are a few minor comments concerning the presentation.

Abstract: ln 15 mentions that species of insect is detected. This is somewhat misleading because the two insects used in this study are from different families. Also, later in ln 19 the term "type of insect" is used. It would be better to use a term such as "taxon" in each case to achieve the correct level of specificity for the technique being used.

Introduction: It would be helpful to have a brief description of the two insects selected, and the rationale for selecting them. Also, the source of the insects and more complete taxonomic information should be provided somewhere in the paper.

Methods: It was unclear if the photos were always performed as they were in Fig. 4. Were there always two individuals of each of these two species with a Canola Seed. Were the arrangements within the field of view random? More details on the photo collection would be helpful.

Figures: Many had font that was difficult to read. Increasing font size and/or switching from a black background would help. Fig 5, 6&8 were pretty difficult for my eyes. Fig. 2 was borderline,

Conclusions: Returning to the question of species identity. This technology is impressive, if it can distinguish two beetles of similar size and shape. However, it should be noted that some species, particularly within genera may never be able to be identified this way because the morphological differences are simply too slight. A brief mention of this might be helpful, with an acknowledgement that this might be simply a good initial screening tool that might require further follow-up. For example, Trogoderma variabile (one of these species), could probably never be distinguished from Trogoderma granarium, and Trogoderma inclusum using this technology.

Author Response

Kindly see attached file for our responses for all three reviewers.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors propose a low-cost and efficient automatic insect detection system. The paper is clearly thought out and the conclusions are basically reliable. Therefore, I recommend accepting this paper after minor revision.

(1) The introduction about machine vision in the introduction is inadequate and its advantages and disadvantages need further expansion.

(2) Figure 5 and Figure 8 do not fit in the academic text; they are mostly presented in tables.

(3) The text in Figure 6 is not presented clearly.

Author Response

Kindly see attached file for our responses for all three reviewers.

Author Response File: Author Response.docx

Reviewer 3 Report

This manuscript introduces an application of machine learning algorithms for insect detection and monitoring, and establishes a comprehensive system encompassing both hardware and software components to automate the monitoring process. This work is a good example to demonstrate the potential and impact of AI on fields such as agriculture and public health. 

The paper is well written. It has very detailed introduction, and the methodology is clearly described. In this dimension the paper is good.

However, the technical novelty is limited. The model this manuscript use is a  very standard one. Although some layers of MobileNet is omitted, there is no innovative changes on the model architecture. Besides that, I am also interested to know why this paper uses MobileNet V1 when V2 and V3 are available and have been proven to have better performance.

The dataset amount seems very small. With 100 test images it is very likely the model can overfit. Is it worthwhile to explore data augmentation on the test set?

In order to improve the manuscript, it would be beneficial to include the following discussions and results:

  1. Performance comparison among various algorithms. The paper mentions that mAP is used to compare performance of different algorithms, but the corresponding results were not presented. Including these results would help to validate the approach and improve the overall quality of the manuscript.

  2. In addition to accuracy, serving latency is an important metric that should be discussed. It would be helpful to have a discussion on the required latency and how it may affect model selection.

  3. I am concerned about how representative the image dataset is. Although the images shown in Figure 4 and 5 are clear, real-world insect detection is more challenging due to insects being in motion and potential camera blurring. A discussion on how this may affect the performance of the model would be helpful to ensure the proposed approach is applicable to real-world scenarios.

Finally, I would suggest that the author consider the possibility of open-sourcing the dataset. This would not only increase the originality of this work, but also contribute to the advancement of the field by providing a valuable resource for other researchers to build upon.

Author Response

Kindly see attached file for our responses for all three reviewers.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

I would like to thank authors to address the comments. The feedback looks good to me.