Real-Time UAV Patrol Technology in Orchard Based on the Swin-T YOLOX Lightweight Model

Round 1

Reviewer 1 Report

The monitoring of diseased plants in orchards with the help of UAVs carrying high-definition sensors is very interesting and challenging. This study proposes a light-weight detection model of Swin-T YOLOX that can serve these purposes well. However, there are several points in this manuscript that need to be revised.

1)      In the abstract section, the discussion of the methodology and model results used in this thesis, the use of tense is not reasonable.

2)      In the fourth paragraph of the Introduction section, the research on "The combination of deep learning and real-time remote sensing by UAV in orchards" is not sufficiently investigated, and references are added.

3)      In the eighth paragraph of the Introduction section, the research on "The model layer pruning technology" is not sufficiently investigated, and the references should be added.

4)      In section 2.2, the authors mention that the data is enhanced, what is the size of the enhanced training set and test set?

5)      The discussion section should be improved. If compared to existing studies, how is this study different and what are the main findings or contributions to the field.

6)      In the conclusion, we need to add the purpose of using the method and summarize the practical application of the model at the end.

Author Response

Dear Reviewer:

Thank you very much for taking time out of your busy schedule to review my paper and give suggestions for revision. I have completed the revision and listed my reasons for revision.

line 18-31, I change the tense of the sentence into the general tense.It's obviously unreasonable for me to use the past tense

line 67-71,I add two references about "Combing deep learning and UAV remote sensing in real-time for orchard" to expound the extensive application of deep learning and real-time UAV remote sensing.

line 109-118, I add seven references about "The model pruning technology".Model pruning technology can be used in CNN and transformer, which shows that the technology is powerful.

line 172-176,I list out the size of the enhanced train set, valid setand test set.

line 433-441,I list out the main contributions of this paper to the patrol orchard field.

line 465-470,I add the purpose of using the method and summarize the practical application of the model, which makes the conclusion more complete.

that's all,

We would like to thank the referee for taking the time to review our manuscript.
--
Best Regards，
Mr. Shaoming Lin

Author Response File: Author Response.docx

Reviewer 2 Report

In their work, the authors demonstrate the use of unmanned aerial vehicles (UAVs) for real-time remote sensing to monitor "diseased" plants or abnormal areas of orchards from low altitudes. This methodology can greatly improve the efficiency and speed of patrol response to vegetation conditions. The aim of the communication of this paper is to demonstrate theoretically and practically the monitoring of an unmanned aircraft and to enable real-time monitoring of an orchard by an unmanned aircraft.
Existing algorithms for small object detection are usually difficult to achieve both in terms of detection accuracy and processing speed. In the present work, a new model for evaluating the desired image, called Swin-T YOLOX , is proposed, which consists of an advanced YOLOX detection network and a powerful Swin Transformer backbone.
In part of the model, layer pruning technology has been applied to reduce the multilayered structure of the Swin Transformer model. A number of strategies were newly proposed and implemented to improve the dataset expansion data in the model training phase. The trained Swin-T YOLOX model was deployed on an embedded Jetson Xavier NX platform to evaluate its performance, detection capabilities of the UAV patrol mission in a real-time environment set.
The research results show that with the help of TensorRT optimization, the proposed lightweight Swin-T YOLOX network achieved 94.0% accuracy and achieved a detection rate of 40 frames per second on the embedded platform.
Compared with the original YOLOX, the accuracy of the model increased by 1.9%, and compared with the original Swin-T YOLOX, the size of the proposed Swin-T YOLOX lightweight network decreased to 1.5%, while the accuracy of the model increased slightly by 0.7%.
The work is based on accepted scientific procedures and contains theoretical suggestions for model design and optimization methods. It demonstrates and compares the newly proposed methods through image experiments.  It contains elements of novelty and is a contribution to the field.
From a formal point of view, I have no major comments.

Author Response

Dear Reviewer:

Thank you very much for your appreciation of our paper!

--
Best Regards，
Mr. Shaoming Lin

Author Response File: Author Response.docx

Reviewer 3 Report

A new lightweight model Swin-T YOLOX is proposed in the manuscript. Combined with the real-time remote sensing technology of UAV, the intelligent terminal of UAV is realized, and the real-time inspection of orchard by UAV is realized, which has achieved good results. However, the following problems need to be further solved:

(1) Introduction: Is the word "cruise" used in line 55 appropriate? "Patrol" is used in the full text except here, which needs to be unified in the full text for the convenience of readers.

(2) Data acquisition and preprocessing: Please properly supplement the parameters of the UAV used and the overall flow chart of the study.

(3) Please supplement the accuracy change caused by the parameter change in Figure 8.

(4) Conclusion and discussion: In the discussion part, the advantages and disadvantages of the model are supplemented with the experimental results.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer:

Thank you very much for taking time out of your busy schedule to review my paper and give suggestions for revision. I have completed the revision and listed my reasons for revision.

line 56,I had check the full text and use "Patrol" uniformly.

line 127,I add the overall flow chart of the study and make the research method clearer.

line 136-137, The UAV here is a data acquisition tool, so the focus is on describing UAV len parameters.

line 287-292,Combining the change of attention heat map in Figure 8 and the feature extraction mechanism of Swin Transformer block, we can explain the accuracy change caused by the parameter change in Figure 9.

line 421-450,I add the method of model lightweight , the main contributions and one shortage of this paper to explain the influence of the advantages and disadvantages of the model on the experimental results.

that's all,

We would like to thank the referee for taking the time to review our manuscript.
--
Best Regards，
Mr. Shaoming Lin

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Dear author:

I think this paper can be published after major revision.  Congratulations!

Best wishes,

Yachun Mao