Identifying Field Crop Diseases Using Transformer-Embedded Convolutional Neural Network

Round 1

Reviewer 1 Report

·       The title is long and not significant

·  In the abstract, authors must present the research problem, the contribution and described briefly the results. You should rewrite it.

·    In this paper you confused between related work and introduction. The introduction part should be the first section and include the motivation of work, it is necessary to mention also the proposed contribution and defined it shortly. Moreover, the rest organization of the paper should be described in this section. The authors should add the introduction section.

·       Rename the literature review to Related Work as section 2.

·    Some sentences are too long to make readers confused, and there are also some typos and grammar errors in this paper.

·       The quality of all figures and tables should be improved.

·       Future works as an integral part should be included in the conclusions.

·  Please improve the reference format and verify the number of each reference cited in the paper

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1795054

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

The authors would like to express their sincere gratitude to the reviewers and editor who gave us many constructive comments and suggestions. We have carefully read the Editor’s and the Reviewers’ comments, and revised the manuscript fully according to the given suggestions. The revisions in the paper are marked by red color. The following is a point-by-point response to their comments:

Responses to the comments of Reviewer #1

Q1:  The title is long and not significant.

A1: Thank you for your suggestions. We have fully considered your suggestions and revised the title of this article. I hope you are satisfied.

 

Q2: In the abstract, authors must present the research problem, the contribution and described briefly the results. You should rewrite it.

A2: Thank you very much for your suggestions. According to your request, we have revised the abstract of the article and revised it in the original paper. (Line 30-46)

 

Q3: In this paper you confused between related work and introduction. The introduction part should be the first section and include the motivation of work, it is necessary to mention also the proposed contribution and defined it shortly. Moreover, the rest organization of the paper should be described in this section. The authors should add the introduction section.

A3: Thank you for your suggestion. We added the introduction part according to your request, and described the organization of the paper in this section. (Line 51-78)

Q4: Rename the literature review to Related Work as section 2.

A4: Thank you for your suggestion. We have seen the writing structure similar to our article in many literatures, such as:

1. Kamal, K.C.; Yin, Z.; Wu, M.; Wu, Z.L. Depthwise separable convolution architectures for plant disease classification. Comput. Electron. Agric. 2019, 165, 104948. https://doi.org/10.1016/j.compag.2019.104948
2. Zhou, J.; Li, J.X.; Wang, C.S.; Wu, H.R.; Zhao, C.J.; Teng, G.F. Crop disease identification and interpretation method based on multimodal deep learning. Comput. Electron. Agric. 2021, 189, 106408. https://doi.org/10.1016/j.compag.2021.106408
3. Picon, A.; Seitz, M.; Alvarez, G.A.; Mohnke, P.; Ortiz, B.A.; Echazarra, J. Crop conditional convolutional neural networks for massive multi-crop plant disease classification over cell phone acquired images taken on real field conditions. Comput. Electron. Agric. 2019, 167, 105093. https://doi.org/10.1016/j.compag.2019.105093

We also saw some articles similar to your description. This may be due to the different writing habits of everyone or the different requirements of journals. But we are pleased to revise the article according to your requirements, so we have divided the literature review into Related Work in Section 2. And the serial numbers of subsequent chapters have been modified. (Line 79)

 

Q5: Some sentences are too long to make readers confused, and there are also some typos and grammar errors in this paper.

A5: Thank you for your advice. We reread the full text and modified some long sentences, spelling and grammar.

 

Q6: The quality of all figures and tables should be improved.

A6: Thank you for your suggestion. Due to our mistake, we forgot to save the picture in HD format, and we have modified it.

 

Q7: Future works as an integral part should be included in the conclusions.

A7: Thank you for your suggestions. We have added the outlook for future work in the conclusion. (Line 537-543)

 

Q8: Please improve the reference format and verify the number of each reference cited in the paper.

A8: Thank you for your suggestion. We have checked each literature and displayed it in the format of this journal. For example:

Wagle, S.A.; Harikrishnan, R.; Ali, S.H.M.; Faseehuddin, M. Classification of plant leaves using new compact convolutional neural network models. Plants 2022, 11, 24. https://doi.org/10.3390/plants11010024

We have corrected the entire reference in this format.

Finally, thank you very much for taking the time to review our paper in your busy schedule, and we are very grateful for this. I hope you are satisfied with our modification. If there are deficiencies, please point out. I would like to express my high respect to you.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors,

I must appreciate the initiative take by you in precision recognition of crop diseases. Further I would like to point out on the following for your clarification.

1. I have marked some points in the manuscript attached, Please respond and justify.

2. Please mention on what basis the background correction was done.

3. A pic may have the effects of the light and shadow. Thus, L ab might be a better option, however, you have done HSV. Please justify.

4. Please mention a paragraph on future scope of research before conclusion which will add the value of the research article. Also, mention if there is any scope for nutrient management based on image analysis and data computation through CNN or ANN. This may open a new arena towards the research in the direction.

5. Overall merit of the article is really good and I believe that you will address my points.

Wish you all the best for engineering the agriculture into the smart direction.

Retards.

Comments for author File: Comments.pdf

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1795054

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

The authors would like to express their sincere gratitude to the reviewers and editor who gave us many constructive comments and suggestions. We have carefully read the Editor’s and the Reviewers’ comments, and revised the manuscript fully according to the given suggestions. The revisions in the paper are marked by red color. The following is a point-by-point response to their comments:

Responses to the comments of Reviewer #2

Q1:  I have marked some points in the manuscript attached, please respond and justify.

A1: Thank you for your suggestion. We found your questions in the article and made corresponding modifications. First place (first row of Table 1): ‘sis’ here should be changed to ‘is’; In the second place (about HSV), we will explain this problem in the third point of this coverletter; The third place: ‘E.g.’ in this paper has been modified as ‘for example’.

 

Q2: Please mention on what basis the background correction was done.

A2: Thank you for your suggestion. We will explain this problem. The acquisition of disease images requires a lot of time and effort from people, and taking the correct disease images requires the photographer to be experienced, which is a problem with using real images for training. Limited by experience, it is difficult for us to obtain the corresponding disease image from the field, so we need to use background replacement technology to generate new data on the basis of existing data. The background replacement technique can simulate the recognition scenarios of different environments by replacing the background of the images based on the existing data, which is efficient and accurate for the whole experimental process.

In fact, background replacement is mainly based on image segmentation technology. It is mainly based on the following operations: The binary image is obtained by binary separation of the target object and the background in the image to be replaced by the background; Processing the interaction area between the target object and the background in the binary image, so that the interaction area becomes a pending area different from the target object and the background, and obtaining a ternary image; Perform background replacement on the image to be replaced based on the ternary image, the image to be replaced and the target background image.

 

Q3: A pic may have the effects of the light and shadow. Thus, Lab might be a better option, however, you have done HSV. Please justify.

A3: Thank you for your suggestion. First of all, the pictures to be replaced are taken under the same light source, so their general environment is similar. Secondly, the background of the picture to be replaced is black. We checked the relevant literature on how to better segment the specified color, and found that HSV plays a greater role in this regard. For details, you can see

Cucchiara, R.; Grana, C.; Piccardi, M.; Prati, A.; Sirotti, S. Improving shadow suppression in moving object detection with HSV color information. Intelligent Transportation Systems. IEEE. https://doi.org/10.1109/ITSC.2001.948679

Finally, after consulting the data, we know that in the shadow detection algorithm, we often need to convert the RGB format image into HSV format. For the shadow area, its chromaticity and saturation change little compared with the original image, mainly because the brightness information changes greatly. By converting the RGB format into HSV format, we can get H, S, V components, so as to get chromaticity, saturation, brightness. In addition, from practical experience, HSV images can obtain binary images with clearer contours. Therefore, we choose to convert RGB format to HSV format for background replacement.

 

Q4: Please mention a paragraph on future scope of research before conclusion which will add the value of the research article. Also, mention if there is any scope for nutrient management based on image analysis and data computation through CNN or ANN. This may open a new arena towards the research in the direction.

A4: Thank you for your suggestion. We have added the outlook of this research in the conclusion of the article. (Line 537-543)

 

Q5: Overall merit of the article is really good and I believe that you will address my points.

A5: Thank you for your encouragement, it is motivating for our future work. We hope that you will be satisfied with our response and that it will inspire our confidence. With all due respect and gratitude to you.

 

Finally, thank you very much for taking the time to review our paper in your busy schedule, and we are very grateful for this. I hope you are satisfied with our modification. If there are deficiencies, please point out. I would like to express my high respect to you.

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have responded satisfactorily to all the reviewer’s concerns. They have made a large number of significant modifications to their paper to improve its quality: adding the suggested modifications in abstract, introduction and related work, improving the paper structure to make it more understandable, improving the results, enhancing conclusion and future works section. 

Reviewer 2 Report

Dear Authors,

It's fine and you have justified properly to my questions. I must appreciate your initiatives for carrying out such research and the present version of the article.

Good luck!

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The paper quality has improved considerably, but I still have comments that should be addressed before my final recommendation. As such, in this review, I will be pointing out those following the order and structure of the paper.

 

Abstract

• Line 22: Dataset1 and Dataset2 provide no information about your solving task. They are just two names without meaning. If you must say that, change it to "in two different datasets, in which the accuracy reached [...] and [...]".
• Line 23: It is not easy to assess some authors' statements. When you say "other excellent models", which will that be? What does excellent mean? That is very vague, no matter how accurate your results were.
• Lines 25-26: You might have reached new ideas from your research, and that's very good for future research, but the output of a paper is not ideas; it is results-based science. Remove this and include actual research conclusions based on your findings of joining Transformer Encoders with CNNs.

 

Introduction

• Line 32: "Relevant studies" is just [1]?
• Line 34: The word "explosive" might not be adequate here.
• Line 38 and 44: It seems to be a hyperlink problem with the citations, which might be related to the template.
• Line 43: Give brief details about the target features.
• Line 55: Do you mean transfer learning? This is a well-known technical term. If you mean so, use the correct terminology.
• Line 69: Is recognition the same as classification? See, classification is the name of the task. If you are working with ML research, regardless of the nature of the research, either theoretical or applied, the terminology is imperative for the reader to understand where your paper stands.
• Line 70: the self-constructed dataset = private datasets.
• Line 73: remove "in series".
• Line 79: there is some difference between performance and efficiency. Performance is better suited to describe the "forecasting power" of a model, and efficiency is more related to the training speed or some tweaks that can make the model efficient, which might not implicate increased performance.
• Line 89: Seems to be a hyperlink problem with the citation.
• Line 94: "Inspired by the Transformer [X] architecture, and more specifically by the Transformer Encoder mechanism [Y], [...]." Where X: https://arxiv.org/abs/1810.04805 and Y: https://doi.org/10.1109/TPAMI.2021.3076155.
• Line 99: I'm missing here, at the end of the Introduction, the traditional paragraph that explains the paper structure and the sections yet to come.

 

Materials and methods

• Line 101: There is a double negative sentence that makes this difficult to understand.
• Line 107: Define complex environment and backgrounds earlier in the paper. I think I do understand what you mean, but someone else might don't.
• Line 116: This is the first I see information about Dataset1 and Dataset2. Here I can understand a little bit more about it. That's why the abstract should be revised.
• Line 126: Break these into a new paragraph. It is hard to follow this way.
• Line 166: When you say about the existing mobile devices, regardless of their status of being limited or not, is where I expect to see some citations about which that could be. Please provide references here.
• Line 185: Here is where the equations start, so please include a table of symbols with all the notations used along with the formulation for easy the reader understanding given the multiple formulations.
• Line 232: What does "hwD" stand for?
• Line 233: Here, it appears differently (Line 232).
• Line 257: It is very clever to consider the different features arising from the disease category and propose an improved loss function to capture that. However, I have to say that this is not a new loss function. It was first introduced in 2016 in the following paper https://link.springer.com/chapter/10.1007/978-3-319-46478-7_31. In good faith, I believe the authors have missed this citation, but I deeply encourage them to go through the whole paper and see if they correctly named (given the correct terminology) and cited (given the original authors) all the techniques they used or they based themselves in. I already approached such a problem in previous comments, and I will skip those from this line forward.
• Line 270: Did you mean local optima instead of optimization?
• Line 272: I understand that efficiency is required when working on an agriculture-related tool to be used in loco with a smartphone. As you are working with images and developing the model on PyTorch, I recommend you to check for APEX training, which is a training routine with mixed-precision using float 16-bit instead of 32-bit for faster convergence and without losing significant performance. Thorugh so, you won't see a difference in performance, but the training and inference time should decrease. Also, in future research, I recommend you to replace Adam with AdamW as it has a better weight decay regularization and works well with amsgrad.
• Line 279: Although you have preprocessed your data with some augmentation techniques in order to have a nearly-balanced dataset, the dataset is not completely balanced. As such, the Accuracy score in Equation 9 should be replaced by the Balanced Accuracy score, which will provide a more reliable view of the reported results. The sensitivity, precision, and F-Score metrics should be replaced by their weighted variant that considers the number of samples in each category for the same reason as before. As such, the reuslts should be recomputed with the correct metrics.

 

Results and discussion

• Lines 297-313 (in red): It is confusing to see the literature review along the results section. Move this to somewhere else, and in Table 2, cite the papers you are referring to explicitly.
• Line 333: What are Plan0 and Plan1? They appear out of the blue. They are not explained or defined. This is not very clear.
• Line 339: Plan4?
• Line 343: You named Plan as a column of the table. Explain this before you set the term and check for the indices you have created because they don't match the writing. Also, what does Param mean? Is that the number of Parameters? What do the values mean and their scale?
• Line 361: Missing space "Figure9(a)".
• Line 398: Define TL and TE and remove the extra blank space.
• Line 404: percentage points? Replace by "%".
• Line 423: Given that you changed the table's name where you stated your model, which is "MOBILENET", do the same in the other tables.
• Line 429: Instead of using "~" use: between X and Y, or something more explicit.
• Line 429: Use achieved instead of gained.

 

References

• Line 521: This seems to be the one reference that does not follow the same pattern for the author's name as the other ones.

Author Response

Please check the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Though few review comments were implemented, majority of the review comments are not addressed. 

Following comments are not addressed by the authors

Comment 2) images from different background can be taken for training and testing

Refer papers “A Deep Learning-Based Approach in Classification and Validation of Tomato Leaf Disease”

“Effect of Data Augmentation in the Classification and Validation of Tomato Plant Disease with Deep Learning Methods”

Comment 3) what is plan4? Mentioned on line 339

Comment 5) what are the nine features in Figure 6 not labelled appropriately.

Comment 9) It would be better if results are shown for cassava and cotton along with apple scab

Comment 10) After updating Table 3 for performance parameter evaluation, the table is having two parts. It's confusing for the reader to understand that table. Make a more clear form of the table that the reader can understand well. Also Table 3 is having results for which plant?

Comment 11) authors can plot the ROC curve with the help of confusion matrix

Comment 13) A comparative analysis of the existing work in the classification of plant diseases with their drawbacks or gaps can be added. Discuss the limitations of the previous works as a motivation for the current study. Author should make a comparative analysis of the table in the introduction section before their proposed work. Authors have discussed only two papers. At Least 8-10 recent papers should be used for  comparative analysis

Comment 14) Formulate the contributions of the authors in this work.

Refer to the paper "Classification of Plant Leaves Using New Compact Convolutional Neural Network Models".

 

Thanks and regards,

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1723756

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

The authors would like to express their sincere gratitude to the reviewers and editor who gave us many constructive comments and suggestions. We have carefully read the Editor’s and the Reviewers’ comments, and revised the manuscript fully according to the given suggestions. The revisions in the paper are marked by red color. The following is a point-by-point response to their comments:

Responses to the comments of Reviewer #2

Comment 2) images from different background can be taken for training and testing

Refer papers “A Deep Learning-Based Approach in Classification and Validation of Tomato Leaf Disease”

“Effect of Data Augmentation in the Classification and Validation of Tomato Plant Disease with Deep Learning Methods”

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. The acquisition of disease images requires a lot of time and effort from people, and taking the correct disease images requires the photographer to be experienced, which is a problem with using real images for training. Limited by experience, it is difficult for us to obtain the corresponding disease image from the field, so we need to use background replacement technology to generate new data on the basis of existing data. The background replacement technique can simulate the recognition scenarios of different environments by replacing the background of the images based on the existing data, which is efficient and accurate for the whole experimental process. In addition, we have read the two articles you recommended and cited them in this article. The major revisions were marked in red (line 580-583).

 

Comment 3) what is plan4? Mentioned on line 339

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. It should be Plan3, we have revised it. The major revisions were marked in red (line 358).

 

Comment 5) what are the nine features in Figure 6 not labelled appropriately.

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. 1~9 in the confusion matrix represent Apple healthy, rust, scab; Cassava bacterial blight, brown streak, healthy, mosaic virus; Cotton boll blight, healthy. The major revisions were marked in red (line 398-399).

 

Comment 10) After updating Table 3 for performance parameter evaluation, the table is having two parts. It's confusing for the reader to understand that table. Make a more clear form of the table that the reader can understand well. Also Table 3 is having results for which plant?

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. Table 3 is explained in more detail. Many papers in CVPR also use this form when conducting ablation experiments. In addition, Table 3 shows the results obtained on Dataset 1, including apple, cassava and cotton.

 

Comment 13) A comparative analysis of the existing work in the classification of plant diseases with their drawbacks or gaps can be added. Discuss the limitations of the previous works as a motivation for the current study. Author should make a comparative analysis of the table in the introduction section before their proposed work. Authors have discussed only two papers. At Least 8-10 recent papers should be used for comparative analysis

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have added more descriptions of the literature on crop leaf disease identification in the introduction. In addition, we have described a lot about the shortcomings of the existing work in the introduction, not only two. Finally, we selected several additional literatures and compared the results in these literatures with this paper, as shown in Section 3.3. In addition, we compare the results of this paper with the results of at least ten papers. The major revisions were marked in red (line 61-75, 456-470).

 

Comment 14) Formulate the contributions of the authors in this work.

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have added the contributions of the authors in this paper. The major revisions were marked in red (line 507-510).

 

Refer to the paper "Classification of Plant Leaves Using New Compact Convolutional Neural Network Models".

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have read the article you recommended and cited it in this article. The major revisions were marked in red (line 520-521)

 

Comment 9) It would be better if results are shown for cassava and cotton along with apple scab

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We are so sorry, we may not be able to satisfy you with the modification of this comment. Although we know that adding relevant data about cassava and cotton in Table 6 will make the article richer and more reliable, it is difficult for me to return to the laboratory to complete the revision of this comment within the time required by the editor due to the impact of COVID-19 and my own health status. I just finished the operation not long ago and am still in the hospital. I am struggling to reply to your suggestions in the hospital bed and try my best to satisfy you. I have sent my hospitalization form to the editor of the journal. You can ask editors. In addition, Dataset 2 is obtained by background replacement technology, which is cumbersome and complex. If all nine diseased leaves applied in this paper are replaced by background, it will be a lot of work.

 

Comment 11) authors can plot the ROC curve with the help of confusion matrix

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We are so sorry again, the modification of this comment may not satisfy you. Like comment 9, the resources around me are limited and I really can't complete the drawing of ROC curve. The sudden COVID-19 and physical discomfort made me unable to prepare sufficient data to complete this modification. The confusion matrix was left by me when I first ran the program in the laboratory. The detailed record of it is not kept on the laptop at hand, so it is difficult for me to draw the ROC curve in this case. Out of a rigorous scientific research attitude, I did not perfunctory this comment. I have done my best to modify the confusion matrix. I sincerely hope you can be satisfied.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

I revised again your manuscript entitled "Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer" submitted to Agriculture journal. All my previous minor comments have been correctly addressed. Unfortunately, the weakest part of the manuscript is still the discussion of results, which has been slightly improved.

 

Major comment:

Section "3. Results and discussion". The discussion of the results must be improved. You need to compare your results to the results from other papers. Scientific work needs a real discussion of results! Additional references are also required.

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1723756

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

The authors would like to express their sincere gratitude to the reviewers and editor who gave us many constructive comments and suggestions. We have carefully read the Editor’s and the Reviewers’ comments, and revised the manuscript fully according to the given suggestions. The revisions in the paper are marked by red color. The following is a point-by-point response to their comments:

Responses to the comments of Reviewer #3

Major comment: Section "3. Results and discussion". The discussion of the results must be improved. You need to compare your results to the results from other papers. Scientific work needs a real discussion of results! Additional references are also required.

Answer: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. Since our datasets is not shared, we found three articles as shown in Table 5 to compare the recognition results respectively. The major revisions were marked in red (line 455-469).

At last, thank you very much for your advices! They have a very important impact on the overall level of this paper! I hope you are satisfied with our answer. If there are other problems in the revised draft, please point out! Thank you again!

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have successfully addressed most of my comments in their new review. However, I'm not satisfied with the questions they did not address. Most of those addressed were purely cosmetic, but the ones essential for the manuscript weren't completed due to one of the authors' time and health condition constraints. Therefore, I recommend that the paper be sent back to the author with enough time for them to recollect the control experiments based on balanced evaluation metrics and provide a new analysis of their findings based on the unbiased results. I'm insisting on this because one cannot claim the model is accurate or efficient without a reliable evaluation. When the dataset labels are unbalanced, the model might be tendencious to one of these classes, and the metrics collected over that model might also be biased towards the majority class. Averaging the results based on the number of elements per class might be an option if no more data can be collected, but that might depend on the results the authors already have and how they were collected.

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1723756

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

The authors would like to express their sincere gratitude to the reviewers and editor who gave us many constructive comments and suggestions. We have carefully read the Editor’s and the Reviewers’ comments, and revised the manuscript fully according to the given suggestions. The revisions in the paper are marked by red color. The following is a point-by-point response to their comments:

Responses to the comments of Reviewer #1

Comment: The authors have successfully addressed most of my comments in their new review. However, I'm not satisfied with the questions they did not address. Most of those addressed were purely cosmetic, but the ones essential for the manuscript weren't completed due to one of the authors' time and health condition constraints. Therefore, I recommend that the paper be sent back to the author with enough time for them to recollect the control experiments based on balanced evaluation metrics and provide a new analysis of their findings based on the unbiased results. I'm insisting on this because one cannot claim the model is accurate or efficient without a reliable evaluation. When the dataset labels are unbalanced, the model might be tendencious to one of these classes, and the metrics collected over that model might also be biased towards the majority class. Averaging the results based on the number of elements per class might be an option if no more data can be collected, but that might depend on the results the authors already have and how they were collected.

 

Answer: First of all, we would like to express our sincere appreciation for your request to the editors for an extension of the deadline. This has allowed the main contributors to this paper sufficient time to recover and respond to your comments in more detail. We have reviewed the data, fully considered your suggestions, and re-established the evaluation metrics, which can be found in Section 2.7 (Line 311). At the same time, we have re-performed the experiments according to the new evaluation indicators, derived the corresponding data, and modified them in the revised draft. We followed your guidance and weighted the obtained model accuracies to address the data imbalance.

Finally, we thank you again for your kindness. We hope you are satisfied with our modification. Best regards to you.

Author Response File: Author Response.docx

Reviewer 2 Report

The review comments are not implemented satisfactorily, irrespective of chance provided for the authors to improve. So the decision is to reject

Author Response

Response to the reviewers’ comments

Manuscript ID:  agriculture-1723756

Manuscript Title:    Recognition of field crop diseases based on Convolutional Neural Network combined with Transformer

By Zhu et al.

         First of all, we apologize for the previous changes that made you dissatisfied. We want you to know that the main contributor of this paper had surgery some time ago and was not discharged until recently, so the previous revision is insufficient. In addition, we communicated with the school, and finally the school allowed us to enter the laboratory to obtain the data we needed. We cherish this opportunity. We have put more energy into this revision and revised every comment you put forward. I hope you can understand our previous situation and re-evaluate our modifications. In addition, we also sent the hospitalization photos of the main contributors to the editors of the magazine, who can prove that we did encounter such trouble in the previous revision process, rather than deliberately not modifying your comments. We tell you this not to emphasize the reason, but we really cherish the opportunity to modify this time. We hope to make the modification to your satisfaction this time. Thank you! Best wishes to you!

Responses to the comments of Reviewer #2

Comment 2) images from different background can be taken for training and testing

Refer papers “A Deep Learning-Based Approach in Classification and Validation of Tomato Leaf Disease”

“Effect of Data Augmentation in the Classification and Validation of Tomato Plant Disease with Deep Learning Methods”

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. The acquisition of disease images requires a lot of time and effort from people, and taking the correct disease images requires the photographer to be experienced, which is a problem with using real images for training. Limited by experience, it is difficult for us to obtain the corresponding disease image from the field, so we need to use background replacement technology to generate new data on the basis of existing data. The background replacement technique can simulate the recognition scenarios of different environments by replacing the background of the images based on the existing data, which is efficient and accurate for the whole experimental process. In addition, we have read the two articles you recommended and cited them in this article. The major revisions were marked in red (Line 586-589).

 

Comment 3) what is plan4? Mentioned on line 339

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. It should be Plan3, we have revised it. The major revisions were marked in red (Line 364).

 

Comment 5) what are the nine features in Figure 6 not labelled appropriately.

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. 1~9 in the confusion matrix represent Apple healthy, rust, scab; Cassava bacterial blight, brown streak, healthy, mosaic virus; Cotton boll blight, healthy. The major revisions were marked in red (Line 405-407).

 

Comment 9) It would be better if results are shown for cassava and cotton along with apple scab

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We redesigned the experimental scheme in Section 3.4, and added cassava brown streak and cotton boll blight to Dataset2. They all carried out background replacement, with 720 pieces in each category, a total of 2,160 pieces. In addition, the data of cassava brown streak and cotton boll blight were added to Table 6, and the data in Table 6 were supplemented and enriched (Line 476-490).

 

Comment 10) After updating Table 3 for performance parameter evaluation, the table is having two parts. It's confusing for the reader to understand that table. Make a more clear form of the table that the reader can understand well. Also Table 3 is having results for which plant?

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have described Table 3 in more detail to ensure that readers can better understand its meaning. In addition, Table 3 shows the results obtained on Dataset 1, including apple, cassava and cotton. These descriptions are supplemented in detail on the title of Table 3 (Line 368-372).

 

Comment 11) authors can plot the ROC curve with the help of confusion matrix

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. The main contributor of this paper has been discharged from the hospital. We communicated with the school and finally obtained the relevant data. The ROC curve has been drawn, and we have added it to Figure 12 and described it (Line 460-464).

 

Comment 13) A comparative analysis of the existing work in the classification of plant diseases with their drawbacks or gaps can be added. Discuss the limitations of the previous works as a motivation for the current study. Author should make a comparative analysis of the table in the introduction section before their proposed work. Authors have discussed only two papers. At Least 8-10 recent papers should be used for comparative analysis

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We added the analysis and comparison of the existing plant disease identification work, elaborated the deficiencies in these studies, and put forward the innovation of our research (Line 75-83,107-119). In addition, we have added a description of the existing work (Line 60-74). Finally, we cited 8 relevant literatures in Table 5 for comparative analysis, their research objects are similar to ours, with high comparability (Line 467-468).

 

Comment 14) Formulate the contributions of the authors in this work.

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have added the contributions of the authors in this paper. The major revisions were marked in red (Line 511-514).

 

Comment 15) Refer to the paper "Classification of Plant Leaves Using New Compact Convolutional Neural Network Models".

A: Many thanks to the expert referee for the precious suggestion, which has been taken seriously. We have read the article you recommended and cited it in this article (Line 524-525).

 

 

We have revised each of your comments, including the parts that we were not able to revise before. This revision has been done with the help of the main contributors to this paper, and we have put more effort into it than before. We believe that this revision is of a higher quality than the previous one. We very much hope that you will give us a chance to re-evaluate our revision. We would really like to receive your approval! Finally, we apologize for the previous shortcomings again. Thank you for reviewing our paper in the midst of your busy schedule. Best wishes to you!

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

The paper has improved significantly from the first round to this one. The authors achieved a sound piece and, therefore, I believe it is ready for publication. A minor comment is related to the references, please double-check if the citations are properly identified. I identified cases where first names are used and last names are abbreviated.

Reviewer 2 Report

After giving multiple chances to improve the content of the manuscript the authors have not justified the reviewer comment implementation.  So the paper is rejected with no resubmission.

Following are the major review comments which could be implemented for improving the quality of the paper. 

Find the comments.  Comment 2: authors said they implemented comment 2 from lines (586-589).. these lines are from the Reference section of the paper. As a reviewer, it's really hard to find the implementation done by authors.     Comment 13: Comparative analysis table would have been a better option. Also here Reference no [24], [25],and [26] are referred. At least 8-10 references are required.   Comment 14: The main contributions of this work mean what the novelty in your paper is and how different is your paper from others. Authors are expected to put 3 to 4 unique contributions to the paper at the end of the Introduction section.