Automatic Detection and Counting of Stacked Eucalypt Timber Using the YOLOv8 Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

• Please highlight your contributions of this paper in the end of Sec. 1
• Please compare the performance for YOLOv8 with other object detection-based methods, such as Faster R-CNN, YOLO-v3, Libra RCNN, etc.
• Please tell the authors how you adjust the hyper-parameters in YOLOv8?
• I am curious about whether the authors consider the data augmentation for data? Also, I think 230 images is not enough to training a model. Could you provide more data?
• I am also interested in the transferability of this method in cross-scene tasks, such as TGRS22-Partial domain adaptation for scene classification from remote sensing imagery. Please add some discussion about the transferability in Sec. 4.
• Some references are recommended to cite in this paper, especially for automatic detection and counting of trees or other objects in forests:
  • RSE23-Surveying coconut trees using high-resolution satellite imagery in remote atolls of the Pacific Ocean (small object detection using SOTA object detection methods)
  • RS23-A New Strategy for Individual Tree Detection and Segmentation from Leaf-on and Leaf-off UAV-LiDAR Point Clouds Based on Automatic Detection of Seed Points (tree detection)
  • RSE23-Surveying coconut trees using high-resolution satellite imagery in remote atolls of the Pacific Ocean (small object detection using SOTA object detection methods)
  • RS23-A New Strategy for Individual Tree Detection and Segmentation from Leaf-on and Leaf-off UAV-LiDAR Point Clouds Based on Automatic Detection of Seed Points (tree detection)
  • RS20-Estimating stem volume in eucalyptus plantations using airborne LiDAR: A comparison of area-and individual tree-based approaches (tree detection)

Author Response

Reviewer 1:

Comment 1:

Please highlight your contributions of this paper in the end of Sec. 1

R: Thank you for your comment. It was placed on line 116.

 

Comment 2:

Please compare the performance for YOLOv8 with other object detection-based methods, such as Faster R-CNN, YOLO-v3, Libra RCNN, etc.

R: Thank you very much for your recommendation. In this work, we have solely focused on the Yolo model in its version 8. Conducting such a comparison is beyond the scope and focus of our study. However, we gratefully acknowledge your suggestion and will consider it for other types of future projects that we are currently developing.

 

Comment 3:

Please tell the authors how you adjust the hyper-parameters in YOLOv8?

R: Thank you very much for your comment. Kindly, we provide you with open access on GitHub, which might address your concern. Please follow the link: https://github.com/ultralytics/ultralytics.

 

Comment 4:

I am curious about whether the authors consider the data augmentation for data? Also, I think 230 images is not enough to training a model. Could you provide more data?

R: Thank you very much for your comment. We understand your concern about the training data; however, the data available is sufficient for the specific industrial sector. It should also be noted that each image contains a certain large number of annotations, this aspect we have placed on line 154.

To make the model operational for large-scale or commercial use, such as developing an app, a greater quantity of data needs to be incorporated. This critical aspect was highlighted in line 360.

 

Comment 5:

I am also interested in the transferability of this method in cross-scene tasks, such as TGRS22-Partial domain adaptation for scene classification from remote sensing imagery. Please add some discussion about the transferability in Sec. 4.

R: Thank you very much for your comment. We have placed it on line 352.

 

Comment 6:

Some references are recommended to cite in this paper, especially for automatic detection and counting of trees or other objects in forests:

RSE23-Surveying coconut trees using high-resolution satellite imagery in remote atolls of the Pacific Ocean (small object detection using SOTA object detection methods)

RS23-A New Strategy for Individual Tree Detection and Segmentation from Leaf-on and Leaf-off UAV-LiDAR Point Clouds Based on Automatic Detection of Seed Points (tree detection)

RSE23-Surveying coconut trees using high-resolution satellite imagery in remote atolls of the Pacific Ocean (small object detection using SOTA object detection methods)

RS23-A New Strategy for Individual Tree Detection and Segmentation from Leaf-on and Leaf-off UAV-LiDAR Point Clouds Based on Automatic Detection of Seed Points (tree detection)

RS20-Estimating stem volume in eucalyptus plantations using airborne LiDAR: A comparison of area-and individual tree-based approaches (tree detection).

R: Thank you very much for your comment. We have kindly considered your recommendations and references.  

Reviewer 2 Report

Comments and Suggestions for Authors

Authors present an approach for automatic counting of stacked timber based on implementation of YOLOv8 model.

Obtained results suggests that this approach could be usefull for a users in the real environments. 
But, commenting from the point of computer scientist, presented work does not have any scientific contribution. It t just presents one possible application of a state-of-the art deep learning model. 
Introduction has some very general sections that shouldn’t be a part of the scientific paper in serious journal although authors may expect that readers might be interested because that info is out if their scope.

Also, description of the image dataset should be much more detailed since it might be of a interest for the readers. Since preparation of a good image dataset for learning of the NN is essential, information on number if kabeled images, their resolution, lighting conditions, used cameras, shooting angle, etc. should be provided.

Now, some additional comments:

in equation (3), with should be corrected to width.

In line 164, authors are mentioning very small objects. How small?
In line 274 and later in conclusions, authors stated that video results were better than imeges. Since this is quite unexpected, explanations should be given. 

Comments on the Quality of English Language

Style should be improved. Native speaker could improve language in general. 

Author Response

Reviewer 2:

Authors present an approach for automatic counting of stacked timber based on implementation of YOLOv8 model.

Comment 1:

Obtained results suggest that this approach could be useful for a user’s in the real environments. 
But, commenting from the point of computer scientist, presented work does not have any scientific contribution. It’s just presenting one possible application of a state-of-the art deep learning model. 

R: Thank you very much for your comment. We appreciate your concern regarding the contribution of this research. We have addressed this aspect in the manuscript, specifically in lines 115 and 371. Additionally, various authors have been applying computer vision methods to different areas within forest management, offering alternative solutions, as cited throughout the manuscript.

Comment 2:

Introduction has some very general sections that shouldn’t be a part of the scientific paper in serious journal although authors may expect that readers might be interested because that info is out if their scope.

R: Thank you very much for your comment. The general versions you mentioned are aimed at readers who are not familiar with the subject, enabling them to understand the research not only from a forestry perspective but also from a computer science standpoint.

Comment 3:

Also, description of the image dataset should be much more detailed since it might be of a interest for the readers. Since preparation of a good image dataset for learning of the NN is essential, information on number if kabeled images, their resolution, lighting conditions, used cameras, shooting angle, etc. should be provided.

R: Thank you very much for your recommendation. We kindly received your suggestion and appreciate your efforts to enhance our manuscript. We have incorporated it on lines 133 and 153.

Comment 4:

Now, some additional comments:

in equation (3), with should be corrected to width.

R: Thank you for your observation. We have corrected it.

Comment 5:

In line 164, authors are mentioning very small objects. How small?

R: Thank you for your comment. We appreciate your kind contribution to improving our manuscript. We have incorporated your recommendation on line 131.

Comment 6:

In line 274 and later in conclusions, authors stated that video results were better than imeges. Since this is quite unexpected, explanations should be given. 

R: Thank you for your comment. We have placed it on line 372.

Comment 7:

Style should be improved. Native speaker could improve language in general. 

R: Thank you for your comment. The manuscript has been reviewed by a professional English language reviewer.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors of the publication presented an analysis of a model for measuring eucalyptus wood in stacks. Round wood, which is the primary raw material in the timber industry, was selected for evaluation . The evaluation of the volume of wood in stacks under forest conditions was carried out using the YOLOv8 model.

 

The abstract correctly presents the methodology, results and basic conclusions of the study. It provides an interesting form of presentation and promotion of the article.

Introduction

The authors described the methods used in the measurement of roundwood.  The basic errors and problems associated with the correct description of the volume of roundwood raw material were pointed out. Methods of manual measurement and directions of automation of measurement and calculation of wood volume were indicated.

Unfortunately, no reference was made to ATRO measurements where mass is used to measure volume. (calculated from density and volume) of dry wood (with 0% moisture content).

Methods of optical assessment with features of both volume measurement and quality characteristics of wood were pointed out. The growing role of AI in research and the possibilities of using neural networks for forest measurements were pointed out.

What was missing was a broader presentation of currently toned scanning methods and programs for calculating timber volume based on the measurement of log fronds with detection of log position (distance from the camera) and consideration of projection.

Methodology

The measurement tool in the form of an iPhone 13 smartphone was described, The measurement object in the form of a stack of eucalyptus wood was presented. Ordered wood and wood with unaligned fronds were separated. Training, test and validation measurement sets were separated. The division of the raw material for testing was correctly and pictorially presented.

Selected learning and validation models for the obtained measurement results were presented.

Results

The developed YOLOv8 model used in the learning process with a high level of precision.  Validated correctly the obtained indicators by image configuration for ordered raw material.

Presented the results of measurements and the error generated for the adopted tests. 

Discussion

The discussion includes information on the development of measurement methods for roundwood. Gradual improvements in the process of detecting objects and scaling and measuring them are indicated. Only learning systems make it possible to achieve an automatic way of modeling and making measurements of raw wood. The reference of measurements with the YELO model for air raids indicates the potential of the studied solution in calculating the volume of cheesewood headed for processing.

The discussion did not address the principles of stack wood measurements and the acceptable errors within stack wood measurements.

Conclusions

Conclusions were formulated in too general terms, lacking the obtained indicators of measurement accuracy for wood in logs. Diameter variation and measurement accuracy for wood of different diameters and variable stacking arrangement were not taken into account.

Additional remarks: The article lacked a literature reference for studies on the measurement of wood in logs for sawmilling and valuable wood for plywood or veneer production. The alignment of the length of the raw material and the alignment of the fronds affect the higher measurement accuracy in the automation of the process. Programs using optical methods for calculating the volume and quality of trees and logs are part of numerous literature studies, so it is worth referring to them.

Author Response

Reviewer 3:

The authors of the publication presented an analysis of a model for measuring eucalyptus wood in stacks. Round wood, which is the primary raw material in the timber industry, was selected for evaluation . The evaluation of the volume of wood in stacks under forest conditions was carried out using the YOLOv8 model.

The abstract correctly presents the methodology, results and basic conclusions of the study. It provides an interesting form of presentation and promotion of the article.

R: Thank you for your comment.

Introduction

The authors described the methods used in the measurement of roundwood.  The basic errors and problems associated with the correct description of the volume of roundwood raw material were pointed out. Methods of manual measurement and directions of automation of measurement and calculation of wood volume were indicated.

Unfortunately, no reference was made to ATRO measurements where mass is used to measure volume. (calculated from density and volume) of dry wood (with 0% moisture content).

Methods of optical assessment with features of both volume measurement and quality characteristics of wood were pointed out. The growing role of AI in research and the possibilities of using neural networks for forest measurements were pointed out.

What was missing was a broader presentation of currently toned scanning methods and programs for calculating timber volume based on the measurement of log fronds with detection of log position (distance from the camera) and consideration of projection.

R: Thank you for your suggestion. It was described in lines 44, 51 and 59.

Methodology

The measurement tool in the form of an iPhone 13 smartphone was described, The measurement object in the form of a stack of eucalyptus wood was presented. Ordered wood and wood with unaligned fronds were separated. Training, test and validation measurement sets were separated. The division of the raw material for testing was correctly and pictorially presented.

Selected learning and validation models for the obtained measurement results were presented.

R: Thank you for your comment.

Results

The developed YOLOv8 model used in the learning process with a high level of precision.  Validated correctly the obtained indicators by image configuration for ordered raw material.

Presented the results of measurements and the error generated for the adopted tests. 

R: Thank you for your comment.

Discussion

The discussion includes information on the development of measurement methods for roundwood. Gradual improvements in the process of detecting objects and scaling and measuring them are indicated. Only learning systems make it possible to achieve an automatic way of modeling and making measurements of raw wood. The reference of measurements with the YELO model for air raids indicates the potential of the studied solution in calculating the volume of cheesewood headed for processing.

The discussion did not address the principles of stack wood measurements and the acceptable errors within stack wood measurements.

R: Thank you for your suggestion. It was described in lines 396 and 414.

 

 

Conclusions

Conclusions were formulated in too general terms, lacking the obtained indicators of measurement accuracy for wood in logs. Diameter variation and measurement accuracy for wood of different diameters and variable stacking arrangement were not taken into account.

R: Thank you for your suggestion. It was described according to our research in line 433.

Additional remarks: The article lacked a literature reference for studies on the measurement of wood in logs for sawmilling and valuable wood for plywood or veneer production. The alignment of the length of the raw material and the alignment of the fronds affect the higher measurement accuracy in the automation of the process. Programs using optical methods for calculating the volume and quality of trees and logs are part of numerous literature studies, so it is worth referring to them.

R: Thank you for your suggestions and comments. We have addressed all of your requests in accordance with our study. This research contributes by presenting a viable application of computer vision technology in the forestry industry, especially for stacked eucalyptus wood for pulp use. To make it easier to identify the changes made to the manuscript, we have highlighted them in red.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed all my issues. 

Author Response

Reviewer 1:

The authors have addressed all my issues. 

R: Thank you so much.

Reviewer 2 Report

Comments and Suggestions for Authors

Authors improved presentation style comparing to the first version of the paper. However, additional explanations that were requested were not addressed properly. Database description is still not properly presented. My question on resolution was aimed towards presenting resolution of the images but also towards the information on size of the stacked bounding boxes since that can be quite important for the network detection results. Also, I have doubts on given in line 135 regarding viewing angle (60deg) since it doesn’t seem right looking at the presented images. Also, information given in line 134 (dpi) is irrelevant for detection process. 
Additionaly, authors should have explained how the images presented in figure 4 were obtained (how the foreground was separated from the bacground that has been blackened).

Finally, I still don’t see enough scientific contribution and also I don’t see who is target audience for the paper. If the forestry researchers are the target, than, in my opinion, very technical info given in the abstract (lines 16-22) is not supposed to be their focus. If the paper is aimed towards the more technical readers related to computer science, than great parts of the introduction are to trivial.

Comments on the Quality of English Language

Language is improved, style could be improved but I don’t have some serious objections.

Author Response

Reviewer 2:

Authors improved presentation style comparing to the first version of the paper. However, additional explanations that were requested were not addressed properly. Database description is still not properly presented. My question on resolution was aimed towards presenting resolution of the images but also towards the information on size of the stacked bounding boxes since that can be quite important for the network detection results.

R: Thank you for your suggestions. The details of the bounding boxes are described in line 170. We have also attached the total number of annotations for your reference.

Also, I have doubts on given in line 135 regarding viewing angle (60deg) since it doesn’t seem right looking at the presented images. Also, information given in line 134 (dpi) is irrelevant for detection process. 
Additionaly, authors should have explained how the images presented in figure 4 were obtained (how the foreground was separated from the bacground that has been blackened).

R: We appreciate your feedback. We confirm that the information described is correct. About the background of Figure 4, it was described on line 278.

Finally, I still don’t see enough scientific contribution and also I don’t see who is target audience for the paper. If the forestry researchers are the target, than, in my opinion, very technical info given in the abstract (lines 16-22) is not supposed to be their focus. If the paper is aimed towards the more technical readers related to computer science, than great parts of the introduction are to trivial.

R: We appreciate your feedback. We understand your concern about the contribution of the work to forestry engineering. We have considered it in the introduction (line 76) and in the discussion (lines 414-426). We believe that the abstract is accessible to forestry engineering professionals with knowledge of computer vision. This research contributes by presenting a viable application of computer vision technology in forestry, aimed at professionals seeking effective solutions in the forestry industry from computer vision.

Author Response File: Author Response.pdf