Deep Learning Application in Plant Stress Imaging: A Review

Round 1

Reviewer 1 Report

The manuscript titled “Deep Learning Application in Plant Stress Imaging: A Review” represents an interesting review. The manuscript is overall well written and easy to read. The manuscript should attract an audience in deep learning, precise agronomy, and phenotyping. The manuscript title is accurate and concise. In the entire manuscript, authors use standard technical and scientific terminology.

After a short Introduction part called Plant stress and sensors, the authors explained the Deep learning principle as well as the potential application of deep learning in plant stress imaging. Research is finished with the Discussion and Outlook sections. I recommended this paper could be reconsidered after major revisions.

Comments for authors:

1. Please expand the abstract with the main findings. Few rows of abstracts are really to low.
2. Please research the more state-of-the-art research on the topic. The introduction section must be expanded by a novel, recent research on the topic of image analysis and image classification in precise agronomy. Suggest to research and include in the manuscript literature about the topic mentioned above as: “Combining computer vision and deep learning to enable ultra-scale aerial phenotyping and precision agriculture: A case study of lettuce production”; “Deep Learning Precision Farming: Tomato Leaf Disease Detection by Transfer Learning”; “An automatic method for weed mapping in oat fields based on UAV imagery”; “Mapping Cynodon Dactylon Infesting Cover Crops with an Automatic Decision Tree-OBIA Procedure and UAV Imagery for Precision Viticulture”.
3. Intensify the discussion, highlight the weaknesses and strengths for some crucial research papers that are explained in this manuscript.
4. Please use the MDPI standard font (Palatino Linotype) on figures if you can. Also, increase the font in figures 1 and increase the size of all figures.
5. The variable names must have the same font style and size in equations, on figures, tables, and in the manuscript text. Please describe/introduce all variables used in equations or on figures in the manuscript text.
6. All equations must be adequately cited in the entire paper.
7. Please, double-check all references and reference style.

Author Response

Review Response: Manuscript # agriengineering-824628

 

We thank the editor and reviewers for taking the time to provide feedback on our manuscript. After reading the editor and reviewer comments, we have carefully modified the manuscript to improve the work based on their suggestions. We provide a more detailed explanation to address the reviewers’ specific comments and the revised content was highlighted with blue color in the manuscript. Furthermore, we revised the format of references citation format based on authors guideline.

 

Comments from Reviewer 1:

The manuscript titled “Deep Learning Application in Plant Stress Imaging: A Review” represents an interesting review. The manuscript is overall well written and easy to read. The manuscript should attract an audience in deep learning, precise agronomy, and phenotyping. The manuscript title is accurate and concise. In the entire manuscript, authors use standard technical and scientific terminology.

After a short Introduction part called Plant stress and sensors, the authors explained the Deep learning principle as well as the potential application of deep learning in plant stress imaging. Research is finished with the Discussion and Outlook sections. I recommended this paper could be reconsidered after major revisions.

Comments for authors:

1. Please expand the abstract with the main findings. Few rows of abstracts are really too low.

Response: Accepted and revised.  The related text was added in abstract ‘We found the deep learning in plant stress has achieved well performance. Also, deep learning in image classification was more widely applied than other image tasks, i.e. object detection and segmentation. While segmentation was one embedded step for image classification and object detection sometimes. (Lines 8-12)

2. Please research the more state-of-the-art research on the topic. The introduction section must be expanded by a novel, recent research on the topic of image analysis and image classification in precise agronomy. Suggest to research and include in the manuscript literature about the topic mentioned above as: “Combining computer vision and deep learning to enable ultra-scale aerial phenotyping and precision agriculture: A case study of lettuce production”; “Deep Learning Precision Farming: Tomato Leaf Disease Detection by Transfer Learning”; “An automatic method for weed mapping in oat fields based on UAV imagery”; “Mapping Cynodon Dactylon Infesting Cover Crops with an Automatic Decision Tree-OBIA Procedure and UAV Imagery for Precision Viticulture”.

Response: Accepted and revised.  The recommended papers were reviewed and cited in this manuscript. (Lines 223-225)

3. Intensify the discussion, highlight the weaknesses and strengths for some crucial research papers that are explained in this manuscript.

Response: Accepted and revised. We added some weaknesses and strengths of algorithms used for segmentation and object detection in section 3. The corresponded text was as follows:

‘R-CNN takes a large amount of train the deep neural network when there are 2000 or more region proposals per image needed to be classified. Meanwhile, there is no learning procedure at first searching stage as the selective search algorithm is fixed. As a result, it may lead to tricky candidate region proposals generated [80].’ (Lines 269-272)

‘R-CNN and Faster R-CNN have been applied to object detection as well, using the regions in the image to locate the object. Recently, YOLO algorithm is often applied for object detection, which uses a single convolutional network to predict the bounding boxes and classify such boxes [85]. YOLO algorithm divides the image into an M × M grids, then m (m<M) bounding boxes are taken within each of the grids. The network yields a class probability for each bounding box. When the bounding boxes have higher class probability than a threshold value, they would be selected and applied for locating the objects in the image. The limitation of YOLO network is that it cannot identify the small objects in the images sometimes [80].’ (Lines 296-303)

4. Please use the MDPI standard font (Palatino Linotype) on figures if you can. Also, increase the font in figures 1 and increase the size of all figures.

Response: Accepted and revised. We tried our best to revise the figures in Palatino Linotype font for all the figures. While some cited figures from original papers cannot be changed, we kept their original font. Also, we revised the figure 1 as recommended.

5. The variable names must have the same font style and size in equations, on figures, tables, and in the manuscript text. Please describe/introduce all variables used in equations or on figures in the manuscript text.

Response: Explained. We checked all the variables in the equations, figures, and tables to double check the variables had the full description. For example, we added note in figure 3 and table 1 to explain the abbreviations.  

6. All equations must be adequately cited in the entire paper.

Response: Accepted and revised. We double checked the instructions for authors, and we revised all equations were in center. Also, we added the cited papers for each equation.

7. Please, double-check all references and reference style.

Response: Accepted and revised. We read thoroughly the manuscript to confirm the cited references in correct format.

Comments from Reviewer 2:

In this paper, the authors provide an introduction describing deep learning and its application to agricultural sector. The paper then focuses on most recent published research, trying to give general indications and comparisons between different crops and approaches.

The topic of the paper is very actual and attracting many research interests.

1. My main concern is related to the originality of the work. I wonder what is the novelty of the paper (besides newer references) compared to the paper: "Deep Learning for Plant Stress Phenotyping: Trends and Future Perspectives", published by Trends in Plant Science, vol. 23/10, 883-898, 2018.

Response: Explained. Singh et al. (2018) reviewed the deep learning application in plant stress detection, which was cited in our manuscript as well. In our manuscript, we reviewed common optical sensors used for plant stress detection. Also, we reviewed different deep learning application purposes in plant stress, including classification, segmentation, and object detection. Furthermore, we reviewed advantage and disadvantages of such above application purposes.   

 

2. By way of example it looks like Table 1 in the mentioned paper has been split in three tables in this new paper.

Response: Explained. The tables in our manuscript aimed to show different deep learning application purposes in plant stress, including classification, segmentation, and object detection, which were different with the table shown in Singh et al. (2018).

3. This paper indeed is a review addressing the same topic with a very similar approach and very similar analyses and conclusions.

Response: Explained. In this manuscript, in sections 4 and 5, we stated the challenges of deep learning in plant stress detection from three perspectives, i.e. imagery, crops, and deep learning technique. Also, we stated our opinions on how deep learning technique combined with optical sensors will improve the plant stress detection. The authors believe our manuscript and Singh et al. (2018) have addressed different issues.

4. I expect the authors answer to this question (how their work is new and original?) and try to improve this paper providing some new insight in the topic.

Response: Explained. As we stated in above questions, Singh et al. (2018) reviewed the deep learning application in plant stress detection, which was cited in our manuscript as well. In our manuscript, we reviewed common optical sensors used for plant stress detection. Also, we reviewed different deep learning application purposes in plant stress, including classification, segmentation, and object detection. Furthermore, we reviewed advantage and disadvantages of such above application purposes.  

Other comments:

5. Abstract is too short and not explanatory/appropriate: please rewrite.

Response: Accepted and revised.  The related text was added in abstract ‘We found the deep learning in plant stress has achieved well performance. Also, deep learning in image classification was more widely applied than other image tasks, i.e. object detection and segmentation. While segmentation was one embedded step for image classification and object detection sometimes. (Lines 8-12)

6. Some meta-analysis on published paper could help better understanding present trends in deep learning applied to agriculture, In tables 1, 2 and 3 "Digital" is very generic: probably digital camera would be more appropriate? or RGB sensor?

Response: Accepted and revised.  The digital in tables 1, 2 and 3 has been changed to RGB sensor.

7. Many references seem out of topic (even though I understand they share the deep learning approach). I would consider removing them:

17 "Mesrabadi..."

28 "Litjens..."

29 "Mahy..."

33 "Kooi..."

80 "Dheeba..."

and other from medical field.

Response: Accepted and revised. Such papers were replaced with papers focused on plant phenotyping.

Author Response File: Author Response.docx

Reviewer 2 Report

Deep Learning Application in Plant Stress Imaging: A Review

In this paper, the authors provide an introduction describing deep learning and its application to agricultural sector. The paper then focuses on most recent published research, trying to give general indications and comparisons between different crops and approaches. 

The topic of the paper is very actual and attracting many research interests. 

My main concern is related to the originality of the work. I wonder what is the novelty of the paper (besides newer references) compared to the paper: "Deep Learning for Plant Stress Phenotyping: Trends and Future Perspectives", published by Trends in PLant Science, vol. 23/10, 883-898, 2018. 

By way of example it looks like Table 1 in the mentioned paper has been splitted in three tables in this new paper. 

This paper indeed is a review addressing the same topic with a very similar approach and very similar analyses and conclusions. 

I expect the authors answer to this question (how their work is new and original?) and try to improve this paper providing some new insgight in the topic. 

Other comments: 

Abstract is too short and not explanatory/appropriate: please rewrite.

Some meta-analysis on published paper could help better understanding present trends in deep learning applied to agriculture,   

In tables 1, 2 and 3 "Digital" is very generic: probably digital camera would be more appropriate? or RGB sensor?

Many references seems out of topic (even though I understand they share the deep learning approach). I would consider removing them:

17 "Mesrabadi..."

28 "Litjens..."

29 "Mahy..."

33 "Kooi..."

80 "Dheeba..."

and other from medical field. 

 

 

Author Response

Review Response: Manuscript # agriengineering-824628

We thank the editor and reviewers for taking the time to provide feedback on our manuscript. After reading the editor and reviewer comments, we have carefully modified the manuscript to improve the work based on their suggestions. We provide a more detailed explanation to address the reviewers’ specific comments and the revised content was highlighted with blue color in the manuscript. Furthermore, we revised the format of references citation format based on authors guideline.

Comments from Reviewer 2:

In this paper, the authors provide an introduction describing deep learning and its application to agricultural sector. The paper then focuses on most recent published research, trying to give general indications and comparisons between different crops and approaches.

The topic of the paper is very actual and attracting many research interests.

1. My main concern is related to the originality of the work. I wonder what is the novelty of the paper (besides newer references) compared to the paper: "Deep Learning for Plant Stress Phenotyping: Trends and Future Perspectives", published by Trends in Plant Science, vol. 23/10, 883-898, 2018.

Response: Explained. Singh et al. (2018) reviewed the deep learning application in plant stress detection, which was cited in our manuscript as well. In our manuscript, we reviewed common optical sensors used for plant stress detection. Also, we reviewed different deep learning application purposes in plant stress, including classification, segmentation, and object detection. Furthermore, we reviewed advantage and disadvantages of such above application purposes.   

 

2. By way of example it looks like Table 1 in the mentioned paper has been split in three tables in this new paper.

Response: Explained. The tables in our manuscript aimed to show different deep learning application purposes in plant stress, including classification, segmentation, and object detection, which were different with the table shown in Singh et al. (2018).

3. This paper indeed is a review addressing the same topic with a very similar approach and very similar analyses and conclusions.

Response: Explained. In this manuscript, in sections 4 and 5, we stated the challenges of deep learning in plant stress detection from three perspectives, i.e. imagery, crops, and deep learning technique. Also, we stated our opinions on how deep learning technique combined with optical sensors will improve the plant stress detection. The authors believe our manuscript and Singh et al. (2018) have addressed different issues.

4. I expect the authors answer to this question (how their work is new and original?) and try to improve this paper providing some new insight in the topic.

Response: Explained. As we stated in above questions, Singh et al. (2018) reviewed the deep learning application in plant stress detection, which was cited in our manuscript as well. In our manuscript, we reviewed common optical sensors used for plant stress detection. Also, we reviewed different deep learning application purposes in plant stress, including classification, segmentation, and object detection. Furthermore, we reviewed advantage and disadvantages of such above application purposes.  

Other comments:

5. Abstract is too short and not explanatory/appropriate: please rewrite.

Response: Accepted and revised.  The related text was added in abstract ‘We found the deep learning in plant stress has achieved well performance. Also, deep learning in image classification was more widely applied than other image tasks, i.e. object detection and segmentation. While segmentation was one embedded step for image classification and object detection sometimes. (Lines 8-12)

6. Some meta-analysis on published paper could help better understanding present trends in deep learning applied to agriculture, In tables 1, 2 and 3 "Digital" is very generic: probably digital camera would be more appropriate? or RGB sensor?

Response: Accepted and revised.  The digital in tables 1, 2 and 3 has been changed to RGB sensor.

7. Many references seem out of topic (even though I understand they share the deep learning approach). I would consider removing them:

17 "Mesrabadi..."

28 "Litjens..."

29 "Mahy..."

33 "Kooi..."

80 "Dheeba..."

and other from medical field.

Response: Accepted and revised. Such papers were replaced with papers focused on plant phenotyping.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have addressed almost all the reviewers' comments, and the manuscript in its current version is improved compared to the original.

I have no further comments, and the revised manuscript can be accepted.

Author Response

Thank you so much for your efforts in improving the quality of our manuscript.

We revised the typos in the manuscript carefully.

Reviewer 2 Report

The paper has not been improved: just some minor amendment has been done (lines 9-11; lines 223-225; lines 270-272; lines 297-303). I believe this is not enough. 

 

Abstract has been integrated with lines 9-11: this has not increased the quality of the abstract (conversely they have many errors). Again: abstract is not suitable and should be rewritten. 

 

My comment: "Some meta-analysis on published paper could help better understanding present trends in deep learning applied to agriculture" has not been addressed. Again: some further analysis would help increasing the quality and originality of the paper. 

 

My comment "I expect the authors answer to this question (how their work is new and original?) and try to improve this paper providing some new insight in the topic." has not been addressed. 

Author Response

Review Response: Manuscript # agriengineering-824628

We thank the editor and reviewers for taking the time to provide feedback on our manuscript. After reading the editor and reviewer comments, we have carefully modified the manuscript to improve the work based on their suggestions. We provide a more detailed explanation to address the reviewers’ specific comments and the revised content was highlighted with blue color in the manuscript. Furthermore, we revised the typos in the manuscript.

Comments from Reviewer:

The paper has not been improved: just some minor amendment has been done (lines 9-11; lines 223-225; lines 270-272; lines 297-303). I believe this is not enough.

1. Abstract has been integrated with lines 9-11: this has not increased the quality of the abstract (conversely, they have many errors). Again: abstract is not suitable and should be rewritten.

Response: Accepted and revised. We rewrote the abstract based on comment as stated in Lines 4-7, 10-15, and 16-18. With the requirement of ‘A single paragraph of about 200 words for Abstract’. We rewrote and controlled the abstract with 195 words.

“Plant stress is one of major issues that cause significant economic loss for growers. While the conventional methods for identifying the stressed plants are labor-intensive. So, it is needed rapid methods for address such issues. With the development of advanced sensing and machine learning techniques,”

“we reviewed different deep learning application purposes in plant stress imaging, including classification, object detection, and segmentation techniques. Image segmentation relates to classification and object detection which should be conducted before segmentation. Meanwhile, both object detection and segmentation tasks are related to classification, plant stress image classification was more widely applied than both these techniques.”

“There are limited open sourced datasets and image labeling is time-consuming. While deep learning as a promising technique will greatly speed up the development of plant stress detection.”

2. My comment: "Some meta-analysis on published paper could help better understanding present trends in deep learning applied to agriculture" has not been addressed. Again: some further analysis would help increasing the quality and originality of the paper.

Response: Accepted and revised. We reviewed more papers and added the detailed methods used for each classification, segmentation, and object detection applications in section 3.1, 3.2 and 3.3 for better understanding present trends in deep learning applied to plant disease diagnosis. (Lines 247-257, 279-286, 325-327). As we aimed to review the deep learning applications in plant diseases detection, the further analysis was not preferred.

“Barbedo, J. G. A. (2019) applied deep learning to detect the individual lesions and spots for 14 plant species [92]. Specifically, this study used a pretrained GoogLeNet CNN for training the models. The images were split into two groups for addressing different objectives. The first group was aimed to image classification, to identify the origin of the observed symptom. While the second one was for object detection which was to identify disease areas amidst healthy tissue and to determine if subsequent classification was conducted or not. The results showed that accuracies obtained using this approach were, in average, 12% higher than those achieved using the original images. The accuracies were higher than 75% for all the considered conditions or number of detected diseases. While the author also claimed that the resized input images for pretrained neural network were not as advantageous as the original images under certain conditions.”

“Lin et al. (2019) applied U-Net CNN to segment and detect the cucumber powdery mildew infected cucumber leaves obtained by RGB sensor [93]. In this study, since the powdery mildew infected pixels were less than that of non-infected pixels, the authors proposed binary cross entropy loss function to magnify the loss value of the powdery mildew infected pixels by 10 times. The results showed that the semantic segmentation CNN model achieved an average pixel accuracy of 96.08% for segmenting the diseased powdery mildew on cucumber leaf images. While it was still challenging to apply such deep neural network in field conditions.”

“R-CNN and Faster R-CNN have been applied to object detection as well, using the regions in the image to locate the object. Recently, YOLO algorithm is often applied for object detection, which uses a single convolutional network to predict the bounding boxes and classify such boxes [85]. YOLO algorithm divides the image into an M × M grids, then m (m<M) bounding boxes are taken within each of the grids. The network yields a class probability for each bounding box. When the bounding boxes have higher class probability than a threshold value, they would be selected and applied for locating the objects in the image. The limitation of YOLO network is that it cannot identify the small objects in the images sometimes [80]. Singh et al. (2020) applied Faster R-CNN with InceptionResnetV2 model and MobileNet model on PlantVillage datasets to detect plant disease which included 2,598 images from 13 plants and over 17 diseases [94].”

3. My comment "I expect the authors answer to this question (how their work is new and original?) and try to improve this paper providing some new insight in the topic." has not been addressed.

Response: Explained. As we stated in Lines 369-373, most of the deep learning applications focused on 2D images and symptomatic plants. We added our insight for future direction in Lines 374-376, “In the future, deep neural networks that can be used for 3D images should be the focus and early detection of the plant disease is pivotal to the precision disease management, especially for that diseases without therapy using pesticide.” Also, we claimed that ‘the semi-supervised and unsupervised deep learning are worthy of being exploratory in the application of plant stress detection, though most of studies are based on the supervised approaches.’ (378-380).

“In the future, deep neural networks that can be used for 3D images should be the focus and early detections of the plant disease is pivotal to the precision disease management, especially for that diseases without therapy using pesticide.”

“Further, the semi-supervised and unsupervised deep learning are worthy of being exploratory in the application of plant stress detection, though most of studies are based on the supervised approaches.”

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

The paper has been properly improved. 

I would recommend an English revision by a native speaker (by way of example the abstract and other parts of the paper have fragmented sentences).