Mapping Gaps in Sugarcane Fields in Unmanned Aerial Vehicle Imagery Using YOLOv5 and ImageJ
, João Pedro Nascimento de Lima 1, Eduardo Antônio Speranza 1 and Fábio Cesar da Silva 1,3Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis study proposed a solution where users can define the size threshold for considering a gap, enabling more precise mapping tailored to the crop's characteristics and growth stage. This method by combining UAV imagery with deep learning techniques, it can help farmers better monitor and manage the defects that occur during sugarcane cultivation and improve productivity and quality. In preliminary tests, the solution using YOLOv5 and ImageJ showed a high success rate, proving the feasibility and effectiveness of the approach. This paper may contain some publishable ideas. But it's also necessary to make some major modifications. Some detailed comments are listed as follows, which will help to improve the quality of the manuscript.
1. The literature review does not effectively drive this study. Why do we need to use YOLOv5 and ImageJ techniques to identify defects in sugarcane fields? The authors need to provide detailed reasons for choosing this model.
2. The use of YOLOv5 and ImageJ techniques to identify defects in sugarcane fields is mentioned in the article, and it is recommended that the working principles and advantages of YOLOv5 and ImageJ be described in detail in 2.3 Section so that readers can better understand the research methodology.
3. In the introduction section, some recent related work on the topics of machine learning and deep learning neural network should be strengthened, several recent investigations are recommended for the author’s references, i.e., " 10.1016/j.compag.2019.104906 "; "10.1016/j.ast.2024.109101 ".
4. Sections 2.1 and 2.2 are too brief and it is recommended that a description of the process of data processing be added.
5. The conclusion section should emphasize the innovative aspects of this study, as well as possible future research directions and applications, and it is recommended that the authors revise the conclusion section to highlight the innovations.
Author Response
Response to Reviewers
Dear Editor,
We appreciate you and the reviewers for your precious time in reviewing our paper and providing valuable comments. It was your valuable and insightful comments that led to possible improvements in the current version. The authors have carefully considered the comments and tried our best to address every one of them. We hope the manuscript after careful revisions meet your high standards. The authors welcome further constructive comments if any. Below we provide the point-by-point responses.
All modifications in the manuscript have been highlighted in yellow.
Sincerely,
Inacio Henrique Yano, PhD
Brazilian Agricultural Research Corporation - Portal Embrapa
Embrapa Digital Agriculture
Centro Paula Souza
Faculdade de Tecnologia de Santana de Parnaiba
Response to Reviewer 1
Comments and Suggestions for Authors
This study proposed a solution where users can define the size threshold for considering a gap, enabling more precise mapping tailored to the crop's characteristics and growth stage. This method by combining UAV imagery with deep learning techniques, it can help farmers better monitor and manage the defects that occur during sugarcane cultivation and improve productivity and quality. In preliminary tests, the solution using YOLOv5 and ImageJ showed a high success rate, proving the feasibility and effectiveness of the approach. This paper may contain some publishable ideas. But it's also necessary to make some major modifications. Some detailed comments are listed as follows, which will help to improve the quality of the manuscript.
Response: Thank you very much
1. The literature review does not effectively drive this study. Why do we need to use YOLOv5 and ImageJ techniques to identify defects in sugarcane fields? The authors need to provide detailed reasons for choosing this model.
Response: Thank you for your comments. We have added the literature review section as reminder.
2. The use of YOLOv5 and ImageJ techniques to identify defects in sugarcane fields is mentioned in the article, and it is recommended that the working principles and advantages of YOLOv5 and ImageJ be described in detail in 2.3 Section so that readers can better understand the research methodology.
Response: The reasons for choosing YOLO were highlighted in the literature review.
3. In the introduction section, some recent related work on the topics of machine learning and deep learning neural network should be strengthened, several recent investigations are recommended for the author’s references, i.e., " 10.1016/j.compag.2019.104906 "; "10.1016/j.ast.2024.109101 ".
Response: Thank you for your comments. The literature review has been included to complement the introduction.
4. Sections 2.1 and 2.2 are too brief and it is recommended that a description of the process of data processing be added.
Response: Thank you for your comments. The two sections have been merged into one and supplemented with new data regarding the images.
5. The conclusion section should emphasize the innovative aspects of this study, as well as possible future research directions and applications, and it is recommended that the authors revise the conclusion section to highlight the innovations.
Response: Thank you for your comments. The aspects related to the innovative solution have been included in the conclusion.
Reviewer 2 Report
Comments and Suggestions for AuthorsSourceURL:file:///home/mymark/Documents/reviews/Mapping gaps in sugarcane fields in UAV imagery through YOLOv5 and ImageJ _rev1.docx
This manuscript presents an interesting work where early detection of sugarcane sprouting gaps that can provide economic advantage when optimizing planting of sugarcane on large-scale farms. However, I have the following comments for the authors to address:
Abstract:
Line 21-24.
The authors report a success rate of 94% without showing how much better it is to the current standard, and also do not show the small gap size metric for which this reported as they say the gap size is determined by the user.
Also, the article focuses on gaps < 50 cm on sugarcane plants between three to five months old, this should be clarified in the abstract.
Introduction:
Line 69-72. The study sets the gap in knowledge as:
Current software like Inforow software has difficulty when gap size < 50 cm when there is occlusion of gap space by sugarcane leaves. However, throughout the manuscript, Inforow results are not presented or compared with the proposed method, and the presented results focus on pixels of 4,200 squared as a measure of gap and not 50 cm, even then the results looks at pixels greater than the 4,200.
Line 72 - 73:
this sentence needs revision to make it clearer: " This problem interferes with smaller
field gaps because the software discards these gaps that seem smaller than 50 cm."
Line 73 - 74: this sentence: "In this work, we used the deep learning neural network YOLOv5 [14] because it is easy to use and also easy to find tutorials available on the Internet", I am of the opinion that ease of use or tutorials should not be a justification for adoption of method but rather ability to solve the problem identified.
Line 75-79: The sentences needs revision to make it clearer. Also, the last sentence: "Later, in a parameterized way and using the ImageJ software, the field gaps that will remain on the map are selected and numbered, i.e., the rural producer is the one who determines the size of the field gap will stay on the map, mitigating the problem caused by the occlusion of sugarcane leaves over the space of the small field gap" may be more appropriate in the Materials and Methods section.
In general, the Introduction section needs major revision, to include literature review on the available methods used to address the problem of gap identification in the presence of occlusion, and also a review of the Yolo and Yolov5 model or references to its architecture. The concluding paragraph should outline what is expected in the remainder of the manuscript.
Materials and Methods:
The section could benefit from a sub-section on Study are and experimental setup, where location, of the study area on a map, as well as tools and specifications for image acquisition and model development and training are organized to make it easier to read. Here, you could also show the fields where drone images are captured for the training and testing of the model.
In addition, Figure 2 could show sample field images from the drone.
Figure 3 could show the Field images and how they are overlapped to generate the orthomosaic image.
Section 2.3 Generate YOLO Model title could be misleading, the authors can use Train YOLOv5 Model to make it more clearer. This is also consistent with Line 167 statement "Upon completion of YOLOv5 training...". Also, the authors can further clarify how images are extracted from the orthomosaic image for annotation and training as depicted in Figure 4.
In my view, Figure 5 Google platform showing a git pull and installation of Yolov5 is generic and does not warrant it being presented as a separate results or process image in the manuscript.
In line 221-225; the author can show or provide reference how the 4,200 pixels^2 selected translate to 50 cm gap. Was there any conversion or calibration procedure involved?
In line 237-240: "The "X" and "Y" columns represent the centroid coordinates of the field gap in the image, which will eventually be replaced with georeferenced data, enabling precise identification and replanting operations in these areas if required.", the manuscript does not provide results of the georeferenced X and Y. The authors need to clarify if this is was done in this research or it is intended for a future work?
In general, the Materials and Methods section should focus on the materials and methods, and results such as Figure 10 and table 1 upwards be presented in the results section.
Also, the authors should clarify why splitting the images into RGB channels are needed if it is just for counting or identification as identification or object IDs can be achieved in Yolo models.
Results:
Lines 321-326: this paragraph may not be needed as the results section focuses on outcomes of the Methods. It may be replaced by training statics such as epochs, batch size, loss functions, convergence criteria, and image loading parameters if it includes any data augmentation, etc.
Lines 339-348: the premise of this research as set forth in the abstract and introduction was to identify occluded gaps less than 50cm. However, the results presented and prioritized has been the identification and annotation of gaps larger than 4,200 pixels^2. This is evident in Figures 15, 16, and 18 using the ImageJ script. This results therefore goes against what the research paper set to addressed.
Discussions:
Lines 373: the authors talk about accuracy, but has not defined the term as this may be ambiguous to other metrics. For instance, the Appendix shows the use of IoU, thus it is not clear what the authors refer to as accuracy in this instance as it has not been defined. The authors should clarify this and discus how this compares to the results of other studies.
Lines 376: the authors reiterates the finding of gaps less than 50 cm (or smaller than 4,200 pixels^2? ) but the results has been focused on greater than 50 cm. The authors clarify
The discussion should also discuss how occlusion by leaves is addressed in this case.
In general, the discussion is poorly written and merely makes claims about potential adjacent applications like pest infestations without addressing how those can be done. Also, it discuss the results of this manuscript with existing literature.
Conclusions:
The conclusion reiterates the identification of gaps smaller than 50 cm but have presented results greater than 4,200 pixels, and have not shown how this compares with existing software (like Inforow as mentioned in the introduction) or provide references to such published results.
Comments on the Quality of English LanguageThe quality of the English language could be improved and the structure of the manuscript revised. The authors should reference the comments for authors for further details.
Author Response
Response to Reviewers
Dear Editor,
We appreciate you and the reviewers for your precious time in reviewing our paper and providing valuable comments. It was your valuable and insightful comments that led to possible improvements in the current version. The authors have carefully considered the comments and tried our best to address every one of them. We hope the manuscript after careful revisions meet your high standards. The authors welcome further constructive comments if any. Below we provide the point-by-point responses.
All modifications in the manuscript have been highlighted in yellow.
Sincerely,
Inacio Henrique Yano, PhD
Brazilian Agricultural Research Corporation - Portal Embrapa
Embrapa Digital Agriculture
Centro Paula Souza
Faculdade de Tecnologia de Santana de Parnaiba
Response to Reviewer 2
Comments and Suggestions for Authors
This manuscript presents an interesting work where early detection of sugarcane sprouting gaps that can provide economic advantage when optimizing planting of sugarcane on large-scale farms. However, I have the following comments for the authors to address:
Response: Thank you very much
Abstract:
Line 21-24.
The authors report a success rate of 94% without showing how much better it is to the current standard, and also do not show the small gap size metric for which this reported as they say the gap size is determined by the user.
Also, the article focuses on gaps < 50 cm on sugarcane plants between three to five months old, this should be clarified in the abstract.
Response: We agree and thank you for your comments. We have revised part of the abstract to clarify the work carried out and the results presented.
Introduction:
Line 69-72. The study sets the gap in knowledge as:
Current software like Inforow software has difficulty when gap size < 50 cm when there is occlusion of gap space by sugarcane leaves. However, throughout the manuscript, Inforow results are not presented or compared with the proposed method, and the presented results focus on pixels of 4,200 squared as a measure of gap and not 50 cm, even then the results looks at pixels greater than the 4,200.
Response: Thank you for your comments. We have rectified the text where Inforow previously struggled to distinguish planting failures ranging from 50 to 100 cm. Additionally, we have clarified how to compare length measurements in centimeters with area in square pixels.
Line 72 - 73:
this sentence needs revision to make it clearer: " This problem interferes with smaller
field gaps because the software discards these gaps that seem smaller than 50 cm."
Response: We acknowledge and appreciate your feedback. This sentence has been relocated to the literature review section for improved clarity.
Line 73 - 74: this sentence: "In this work, we used the deep learning neural network YOLOv5 [14] because it is easy to use and also easy to find tutorials available on the Internet", I am of the opinion that ease of use or tutorials should not be a justification for adoption of method but rather ability to solve the problem identified.
Response: We appreciate your feedback and have revised our rationale for selecting YOLO, providing a more detailed explanation in the literature review.
Line 75-79: The sentences needs revision to make it clearer. Also, the last sentence: "Later, in a parameterized way and using the ImageJ software, the field gaps that will remain on the map are selected and numbered, i.e., the rural producer is the one who determines the size of the field gap will stay on the map, mitigating the problem caused by the occlusion of sugarcane leaves over the space of the small field gap" may be more appropriate in the Materials and Methods section.
Response: Thank you for your feedback. We have condensed this paragraph to better suit the introduction.
In general, the Introduction section needs major revision, to include literature review on the available methods used to address the problem of gap identification in the presence of occlusion, and also a review of the Yolo and Yolov5 model or references to its architecture. The concluding paragraph should outline what is expected in the remainder of the manuscript.
Response: Thank you for your feedback. We have added a literature review section to complement the introduction.
Materials and Methods:
The section could benefit from a sub-section on Study are and experimental setup, where location, of the study area on a map, as well as tools and specifications for image acquisition and model development and training are organized to make it easier to read. Here, you could also show the fields where drone images are captured for the training and testing of the model.
Response: We appreciate your comments. The location, image capture process, and orthomosaic construction procedures have been included.
In addition, Figure 2 could show sample field images from the drone.
Figure 3 could show the Field images and how they are overlapped to generate the orthomosaic image.
Response: Thank you for your feedback. The images were overlapped by 90%. Pix4D software facilitates drone flight planning and includes a parameter for overlapping percentage.
Section 2.3 Generate YOLO Model title could be misleading, the authors can use Train YOLOv5 Model to make it more clearer. This is also consistent with Line 167 statement "Upon completion of YOLOv5 training...". Also, the authors can further clarify how images are extracted from the orthomosaic image for annotation and training as depicted in Figure 4.
Response: We appreciate your comments. The text has been revised according to your suggestions.
In my view, Figure 5 Google platform showing a git pull and installation of Yolov5 is generic and does not warrant it being presented as a separate results or process image in the manuscript.
Response: Thank you for your feedback. The image depicting the training process has replaced the previous one.
In line 221-225; the author can show or provide reference how the 4,200 pixels^2 selected translate to 50 cm gap. Was there any conversion or calibration procedure involved?
Response: We appreciate your comments. A new section has been included between lines 262 and 324 to explain the measurement methods used.
In line 237-240: "The "X" and "Y" columns represent the centroid coordinates of the field gap in the image, which will eventually be replaced with georeferenced data, enabling precise identification and replanting operations in these areas if required.", the manuscript does not provide results of the georeferenced X and Y. The authors need to clarify if this is was done in this research or it is intended for a future work?
Response: Thank you for your feedback. We have corrected the text and noted that it will be addressed in future work if requested by rural producers.
In general, the Materials and Methods section should focus on the materials and methods, and results such as Figure 10 and table 1 upwards be presented in the results section.
Response: We acknowledge your feedback. This small image and table with a few lines serve as examples to clarify the procedure, rather than presenting results.
Also, the authors should clarify why splitting the images into RGB channels are needed if it is just for counting or identification as identification or object IDs can be achieved in Yolo models.
Response: Thank you for your comments. The majority of calculations and transformations performed by ImageJ are on 8-bit images, each with only one channel. This necessitates dividing them into channels.
Results:
Lines 321-326: this paragraph may not be needed as the results section focuses on outcomes of the Methods. It may be replaced by training statics such as epochs, batch size, loss functions, convergence criteria, and image loading parameters if it includes any data augmentation, etc.
Response: We appreciate your feedback. The text has been revised in accordance with your comments.
Lines 339-348: the premise of this research as set forth in the abstract and introduction was to identify occluded gaps less than 50cm. However, the results presented and prioritized has been the identification and annotation of gaps larger than 4,200 pixels^2. This is evident in Figures 15, 16, and 18 using the ImageJ script. This results therefore goes against what the research paper set to addressed.
Response: Thank you for your feedback. Explanations on accuracy assessment have been included between lines 460 and 471.
Discussions:
Lines 373: the authors talk about accuracy, but has not defined the term as this may be ambiguous to other metrics. For instance, the Appendix shows the use of IoU, thus it is not clear what the authors refer to as accuracy in this instance as it has not been defined. The authors should clarify this and discus how this compares to the results of other studies.
Response: We agree and thank you for the comments. Explanations on obtaining accuracy were included between lines 460 and 471.
Lines 376: the authors reiterates the finding of gaps less than 50 cm (or smaller than 4,200 pixels^2? ) but the results has been focused on greater than 50 cm. The authors clarify
Response: We acknowledge your feedback. A new section has been added between lines 262 and 324 to elaborate on the measurement methods used.
The discussion should also discuss how occlusion by leaves is addressed in this case.
Response: Thank you for your comments. We have removed this part as it will be considered for future implementation.
In general, the discussion is poorly written and merely makes claims about potential adjacent applications like pest infestations without addressing how those can be done. Also, it discuss the results of this manuscript with existing literature.
Response: We appreciate your feedback. We have excluded untested possibilities from this work and consolidated the Discussions and Conclusions sections.
Conclusions:
The conclusion reiterates the identification of gaps smaller than 50 cm but have presented results greater than 4,200 pixels, and have not shown how this compares with existing software (like Inforow as mentioned in the introduction) or provide references to such published results.
Response: Thank you for your feedback. We have rewritten this section to provide a clearer explanation of the measures used.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper is organized fine.
