HHS-RT-DETR: A Method for the Detection of Citrus Greening Disease

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper is devoted to the topical topic of timely and accurate detection of fruit disease, which causes significant damage to the agricultural industry. The authors propose a method for identifying huanglongbing disease based on the application of a deep learning model. To train the model, a proprietary data set was collected, expanded using augmentation methods. In addition, experiments were conducted to analyze the effectiveness of the proposed approach using the PASCAL VOC dataset.

The paper describes the proposed approach in sufficient detail. The experimental results confirm its effectiveness according to the criteria of accuracy and the number of parameters. The approach is based on new results in the field of deep learning. The list of references contains relatively new works and corresponds to the topic of the paper.

The paper as a whole makes a positive impression. However, almost before the conclusion, it is unclear why it is necessary to reduce the number of model parameters and why comparison is carried out only with lightweight versions of the YOLO model. Only at the very end it is said that it is planned to implement the proposed approach in unmanned aerial vehicles. That explains a lot, since unmanned aerial vehicles, as a rule, use computing devices with limited resources. But it can be recommended to outline the conditions for applying the results of the work early, since this is associated with the formulation of a research task.

This remark is of a recommendatory nature and does not reduce the overall positive assessment of the work. 

Author Response

Dear Reviewer,

    I have made revisions in accordance with your questions and suggestions, and I have indicated the locations of the changes in the attachment.Thank you very much for your suggestion, which has greatly improved the quality of my paper.Please see the attachment.

Best wishes,

Yi Huangfu

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. the authors should include the conclusion and future work in the abstract as well.

2. There is not connectivity of the method with the experimental results in the abstract.

3. How to authors validate the statistical results of recall rate and the average accuracy in the abstract.

4. The references are not cited properly and not in order. The must focused on the specific reference with properly cited.

5. The must include the problems contribution at the end of the introduction.

6. The authors must include the literature review related to the topic

7. The authors did not provide the source of the Figure 1 and Table 1

8. How the authors implement the experimental materials used in the manuscript.

9.  How the authors used RT-DETR-r18 model structure (Original model structure)?

10. What is the relation of Figures 3 - Figure 9 with the title?

11. Figure 11 is the just the detection but how the authors contribute his work.

12. Why the used Figure 12 for different models.

13. How the authors validate their work.

14. The conclusion must rewrite and compare with the abstract objective.

15. The references must be update and cited properly.

Author Response

Dear Reviewer,

   I have made revisions in accordance with your questions and suggestions, and we have indicated the locations of the changes in the attachment.Thank you very much for your suggestion, which has greatly improved the quality of my paper.Please see the attachment.

Best wishes,

Yi Huangfu

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper introduces a database with images containing three types of oranges that may be affected by Citrus greening disease. The paper than establishes a baseline using a  RT-DETR-r18 type of deep network. Then the model is evolved into a so-called HHS-RT-DETR

Unfortunately, in the current form the paper does not have enough contribution to claim "acceptance".

Issues:

1. (language) The paper is full of Chinese names while the paper is English. For instance, "Huanglongbing" is the Chinese name. The English name is "Citrus greening disease/greening". "Wokan", "Daejong", etc.

2. (observation - not issue) The main contribution of the paper seem to be the introduction of a new database with establishing a baseline performance using a predefined model. Therefore the agronomy contribution is w.r.t database, while the machine learning is more towards improvement from RT-DETR-r18 to HHS-RT-DETR.

3. Database-agronomy: using only three types of oranges is not impressive at all. Furthermore, let me quote from ref [1]:

"Greening (huanglongbing) is the most serious disease of citrus that limits production in the subtropical and tropical citrus-producing areas of the world (Gottwald 2010). Before its identification, the disease was known by a variety of names: yellow shoot (huanglongbing) in China; likubin (decline) in Taiwan; dieback in India; leaf mottle in the Philippines; vein phloem degeneration in
Indonesia; and yellow branch, blotchy mottle, or greening in South Africa. As it became clear that all these diseases were similar, the term “greening” was widely adopted (da Graça 1991). The disease is caused by three phloem-limited bacteria: “Candidatus Liberibacter asiaticus” (CLas), “C. Liberibacter africanus” (CLaf), and “C. Liberibacter americanus” (CLam) (Bové 2006, Gottwald 2010). CLas and CLam are predominantly transmitted by the Asian citrus psyllid (ACP; Diaphorina citri Kuwayama), whereas CLaf is transmitted by the African citrus psyllid Trioza erytreae Del Guercio."

The information presented in this paper is significantly different, and I would say inaccurate.

4. Database - machine learning perspective:

a. Size: having 1200 of different images is, again, not impressive. The extension by augmentation is automatically done in any deep learning code. Furthermore since the database was manually extended, there is a significant possibility that an original image and its extensions to fall in both training and testing which make the ML problem ridiculously easy.

b. Annotation: the disease may be seen on individual leaves (BTW, this is not specified and until, at page 8/figure 9, I believed that is about fruits) but somehow the annotation are with bounding box including a package of leaves affected by the disease. There is no motivation to mark a block of leaves and, especially since it contains both ill leaves and healthy leaves, there is no reason to focus on predicting the geometrical coordinates of the box accurately. The point is: given this type of annotation, better localization is not valuable.

5. Method: the beginning of the method is standard and the paper uses large chunks to shuffle through known things (e.g. RT-DETR-r18). The point here is: either the paper assumes a reader familiar with the model , procedure case in which section 3 should be significantly reduced, either the paper assumes that the reader is not familiar with the machine learning aspects of this model and it needs to explain everything. The more likely is the second choice, but than the presentation lacks significant details. For instance:
- what is the meaning of all blocks from figure 3/4? What does it mean distance^{sℎape} in eq. (3) (this is important since it is a claim of novelty)?

-why putting in parallel maxpooling and average pooling should help?

- there is a significant problem with figure 10: maxpooling is always larger or at most equal to average pooling. Therefore image (b) should be noticeable brighter than image (c). Now is viceversa.

- Also what is in the HWD filter that the output image is brighter than the input? HWD is a concatenation of selected band-pass frequencies from the input without any amplification. It is more likely that the output of the HWD filter is the one from figure 9 (which is totally different from the one in figure 10).

6. Evaluation: it is not clear.

a.The key metric is mAP, which is based on AP(i) detailed in table 3. There it is an integral over Recall values from 0 to 1. How is Recall varied continuously? Shouldn't be a sum over a quantized set of Recalls? If this is the case what is varied to get different Recall values? Confidence threshold?

b. Table 4: why by introducing two components, the number of parameters decrease?

c. Section 4.3 . Now it is the first time when GradCam is mentioned. Where is introduced? Assuming that it is a module added to visualize the activation of some final layer, there is no reason to have the discontinuity from the bounding box (figure 13)

d. Table 6 - the values for accuracy and params for Yolo seems to be wrong. Please take a look, for instance at table 1 from "YOLO-v1 to YOLO-v8, the Rise of YOLO and Its Complementary Nature toward Digital Manufacturing and Industrial Defect Detection" https://www.mdpi.com/2075-1702/11/7/677

e. Table 7- this is just bad. Yolo v1 reaches 63.4 mAP as reported  in the original paper (see table 1 from original paper) if it is PASCAL VOC 2007 and 57.9 if it is VOC 2012. Reports around 30 mAP means that there is a significant error. Unfortunately this casts a significant shadow over the entire implementation.

Comments on the Quality of English Language

Please see point 1 from previous sections

Author Response

Dear Reviwer,

   I have made revisions in accordance with your questions and suggestions, and I have indicated the locations of the changes in the attachment.Thank you very much for your suggestion, which has greatly improved the quality of my paper.Please see the attachment.

Best wishes,

Yi Huangfu

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

There are a lot of misunderstandings between Figure 3 to Figure 8. It is important for the authors to focus on their work. It is recommended that the authors only include the modified model, and any other models cited in the text

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Dear Reviewer，

      Thank you very much for your feedback. We have carefully considered your suggestion this time. We do have such issues in our paper. We have made the corresponding modifications.Please see the attachment.

Best wishes，

Yi Huangfu

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The reviewed version greatly improve the manuscript. While there has not been any improvement of the method, it improve the explanation and argumentation.

In the current form I view this paper as passable, even that is by a slight margin. The innovation is limited and database too, but there may be an auditorium interested in it.

 

What could be improved is the formatting: currently there are tables that are spread over two page, figures should be either full page wide or full text wide, but not both, big blank on page 6. Anyway, these should be checked with editorial office.

Author Response

Dear Reviewer，

   Thank you very much for your feedback. We have carefully considered your suggestion. Please see the attachment.

Best wishes，

Yi Huangfu

Author Response File: Author Response.pdf