Accurate Detection Algorithm of Citrus Psyllid Using the YOLOv5s-BC Model

Round 1

Reviewer 1 Report

The manuscript contains original results on the detection of citrus psyllid using the YOLOv5s-BC model.

High recognition accuracy was obtained. Therefore, the developed procedure can be useful in practice. However, several issues need to be improved.

Symptoms of infection should be described in more detail in the Introduction.

3.1 Experimental Data Collection and Data Set Construction: Were the images acquired every day?

line 251: How many images were in each group?

The obtained results should be explained and discussed in more detail.

The Discussion should be expanded and include references.

Directions for further research should be indicated in more detail.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents a method for detection of citrus psyllids from images collected using mobile devices. A state-of-the-art YOLOv5s model was trained and improved by adding additional convolutional layers. Despite the attempt, it was found that the improvement was quite minimal; this was concluded since the YOLOv5s model without improvement also performed well but some other parameters may have not been optimized. It was observed that certain solutions such as by changing the network from YOLOv5s to YOLOv5m/l or other models were not tested and may actually improve the performance, better than the current approach. Otherwise, other techniques such as object level augmentation or tiling may also help in improving the detection results. In an agronomy point of view, the method for properly acquiring the images was also not reported, which is rather the most important aspect in this context. Therefore, this publication is not recommended for publication. 

Major comments:

    1. L107: What was the main reason that YOLOv5s was preferred over its other variants such as YOLOv5m, etc? Doesn’t that also affect the performance of the model?

    2. L189-191: What was the basis in putting the two SENet layers in the 8th and 10th layers? Will this be the same if YOLOv5m/l was used?

    3. L220: Isn’t this the mosaic data augmentation embedded in YOLOv5 training? If yes, what was the probability set for augmentation?

    4. Figure 6: What if there were no psyllids in the mosaicked image?

    5. L244-250: Given this image acquisition setup, how will this be applied in real practice?

    6. L250: How were the resolutions of 900x900 and 600x600 selected? Does that mean that the images were resized/rescaled for training? This might be crucial since other mobile devices have really high/good image resolutions but it wasn’t fully utilized.

    7. L251: What were the percentage of these groups?

    8. Figure 8: From these two graphs, it is quite obvious that there wasn’t a significant improvement by adding CBAM or SENet layers. It clearly shows that the YOLOv5s was sufficient already for this application. Besides that, the results in Figure 10 was also not quite convincing enough. Was the NMS threshold optimized? How were these sample images also selected?

 

Other technical questions:

1. Were the anchor boxes tuned? Since very small objects were to be detected, anchor boxes must be adjusted to fit the target objects.

2. Where do the authors expect to process the images? Mobile device or via server? If via server, then the processing speed may not be an issue and YOLOv5m/l may be better options.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Huanglongbing (HLB), also known as citrus greening, is a highly destructive and lethal bacterial disease that affects citrus trees. Detection of HLB can be challenging as symptoms often take months or years to appear and may resemble other citrus diseases. In principle, there are three methods of detection: visual inspection, laboratory testing, trapping and detection of psyllids. But all three analyses are complex and time consuming.

Early detection and rapid response are crucial in controlling the spread of HLB and protecting citrus trees. Currently, there is no cure for HLB, and the best method of control is to prevent the spread of the psyllid vector through quarantine measures and aggressive management practices. Considering these facts, the aim of this paper is represented by the detection of the small-sized citrus psyllid based on the YOLOv5s algorithm in terms of target detection, combined with an SE-Net attention module.

The introduction section present the importance of development of accurate detection technology for this pathogen that will contribute to controlling and preventing the transmission of HLB. Authors present the last solutions based on CNN that have been proposed to accurately identify orchard diseases and pests in the natural environment. Based on the findings the authors propose a detection algorithm for small targets based on BottleneckCSP and YOLOv5s-BC.

Section 2 present clear and to the point the analysis and design of YOLOv5c-BC algorithm.

In section 3, Experimental Design and Analysis, the steps for collecting the experimental data set are explained. 

In conclusion, the paper investigated the effectiveness of the improved YOLOv5s-BC algorithm. It was shown that YOLOv5s-BC algorithm exhibits improved accuracy and reduced missed detection of citrus psyllid with 2.41% higher than that of traditional YOLOv5s, but with the mention that the algorithm must be further developed.

Few remarks before to accept this paper:

- At figure 9 - on the ordinate (y axis) there are some Chinese notes. Please correct the figure.

- It will be better if you will have Table 6 on the same page. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The manuscript has been significantly improved.

Reviewer 2 Report

I believe the revisions and the replies to the review comments made the manuscript better.