Multi-Crop Navigation Line Extraction Based on Improved YOLO-v8 and Threshold-DBSCAN under Complex Agricultural Environments

Round 1

Reviewer 1 Report

The article is devoted to the construction of the navigation line of planted crops (including cabbage, kohlrabi and rice) of both straight and curved trajectories, by combining the improved YOLOv8 neural network and the DBSCAN clustering algorithm algorithm.

The article is of scientific and practical interest.

Notes on the article.

1) At the end of section 1, clearly formulate the purpose of the study and tasks: 1,2,3...

2) It is also advisable to bring the conclusions into line with the objectives.

3) Some of the conclusions can be transferred to the “Discussion” section, where it is also necessary to compare your method with the studies of other authors.

4) Fig. 11-14 - Specify which crops are shown, also sign parts of the pictures (a, b, c...) with explanations

5) Research methodology should be improved, perhaps using a mind mapping tool. Alternatively, provide a graphical summary, this will allow the reader to immediately understand the structure of your research.

6) It would be good to provide a fragment of the dataset with different versions of photographs (from a camera, from a drone - what is described in section 2.1), for example, 20-30 small photos.

7) Also focus on the practical significance of the research, how it will allow agricultural machinery to choose the right route, how this can be practically implemented, what economic effect is expected due to your research.

In general, the article makes a positive impression, is well designed, and can be published in the journal "Agriculture" taking into account the comments.

Author Response

please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors, the article is well written and comprehensible. I just address you few notes
- lines should be numbered to allow a punctual review
- the audience of Agriculture are not only computer Scientists, so pay attention to the used language - e.g.In several points you address 'other crops' (e.g. discussing the white boxes in the last figures) which should be addressed as "weeds".
- though the number of picture is large enough the cropping context is not described with sufficient detail.
- Results do not evidence enough a comparison with other authors' results
 
- Application notes should be added as well, answering to questions as:
-- how complex is the (re)training in a completely new environment ?
-- which is the power required for an on-board computer to run your algorithms real-time on a tractor ?
-- which is the sensitivity of your approach to the angle view of cameras ?
- in general which is the robustness of your method ?
 
 - Fig 9,10, 11,12, 13, 14 - please increase readability

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study proposes a combined multi-crop row centerline extraction algorithm based on improved YOLOv8 model, threshold DBSCAN clustering, least squares method, and B-spline curves.

The introduction section also provides an in-depth analysis of related work for navigational line extraction in various crops and with various approaches. Most references are recent and relevant to the subject discussed, but the section does not mention any metrics obtained (accuracy etc..) which would be useful for later numerical comparison.

A dataset of images is also collected, then augmented for variety in model training.

The DCGA-YOLOv8 model is described with the additional layers for improvement, to detect row of multiple crops, and several performance indicators are used for evaluation.

The results for the F1-score and mAP values for Cabbage, Kohlrabi, and Rice are 96.4%, 97.1%, 95.9%, and 98.9%, 99.2%, 99.1%, respectively. A threshold-DBSCAN algorithm is proposed for each row, with the correct clustering rate reaching 98.9%, 97.9%, and 100%. LSM and cubic B-spline curve methods are applied for straight and curved crop rows.

The method improves the safety and stability of visual navigation and field operation of agricultural machines, compared with YOLOv8, YOLO-v5, Faster R-CNN and SSD model.

The paper is well written with very interesting results, but it lacks some discussion points in terms of limitations of the work:

·        - It would be interesting to compare the proposed model with other similar approaches (not just mainstream models) to assess how the results are compared with the state of the art presented in section 1.

·        - Also as part of discussion, what would be the behavior of the model with a more noisy input (images with bad lighting conditions, shadow/sun exposure, bad resolution, etc. In other words, which of these do matter for the model to perform well and which are limitation of the model and require further work? The problem of the steep angle is discussed, but how about the others?

·        - What would the model require to work with more types of crops? In this case we talk about multi-crop, but there is a limited and very reduced number of types of crops involved actually. What would mean to generalize this to 10-20 types of crops?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

In their responses, the authors slightly corrected the manuscript, signed parts of the figures, added prospects for future research, and added several photos of a training set of photographs. However, many of my comments remained uncorrected or the answers were formal.

 I still recommend presenting a research diagram in the form of a drawing - either a mind map or a graphic annotation.

The signed parts of pictures a, b, c blend into the dark background, highlight them with a bright color.

I would recommend making a separate “Discussion” section (I leave it to the editors’ discretion).

Present your research objectives more clearly: 1,2,3….

Bring your conclusions in line with them.

I also asked to present more specifically and in detail the practical meaning of the study. What will be the impact of retrieving agricultural machinery navigation lines? To reduce crop losses, to reduce fuel consumption, to speed up harvesting? What is the economic effect (in quantitative units), due to what? This is very important, since the article will be published in the journal “Agriculture”.

The article is generally good, but I would like to once again see real adjustments in response to my comments, rather than formal ones.

Author Response

Response to Reviewer 1 Comments

We apologize for the careless revision the first time, and thank you for giving us the opportunity to make a second revision. We would like to thank you for your valuable comments, which will greatly improve the quality of this manuscript.

1. I still recommend presenting a research diagram in the form of a drawing - either a mind map or a graphic annotation.

Responses:

According to your suggestion, we have added a research flowchart in the revised manuscript.

The signed parts of pictures a, b, c blend into the dark background, highlight them with a bright color.

Responses:

We highlighted it in the revised manuscript.

2. I would recommend making a separate “Discussion” section (I leave it to the editors’ discretion).

Responses:

According to your suggestion, we have added “Discussion” section in the revised manuscript.

Discussion

The detection of multiple crops in agricultural environments often suffers from robustness and generalization problems due to differences in the type, shape, and size of crops. In this study, an improved YOLO-v8 and threshold-DBSCAN algorithm was proposed to improve the performance of multi-crop detection and crop row fitting.

In terms of multi-crop detection, the DCGA-YOLOv8 model proposed in this study shown superior detection accuracy and robustness, compared to other mainstream recognition models. The DCGA-YOLOv8 model pays more attention to the important feature areas of crops by adding global attention to suppress useless or invalid features, and introduces deformable convolution in the detection process to obtain more fine-grained spatial information between crops, thereby improving the accuracy of crop detection. However, the DCGA-YOLOv8 model still has some difficulties in crop detection at image boundaries. The unclear boundaries of crops and the partial overlap between crops make it difficult to accurately label the complete area of a single crop when performing manual labeling. Therefore, the intersection ratio between the prediction box and the ground truth will be relatively small in the detection process, which will lead to detection failure. Nevertheless, DCGA-YOLOv8 has a very high recognition accuracy, and the centroid of the adjacent crop prediction box can still reflect the trend between crops, without affecting the fitting of the subsequent crop row centerline.

For clustering of crops, the distance threshold detection was introduced to exclude interfering crops before DBSCAN clustering. A high clustering accuracy was achieved for each crop, especially for rice, where the clustering accuracy was 100%. It is worth noting that cabbage and Kohlrabi still have crop clustering errors. This may be due to the fact that there are fuzzy boundaries or overlaps between crops in some images, and it may be difficult for the DBSCAN algorithm to accurately distinguish them. To solve this problem, the change in angle between crop centroids is taken into account in the subsequent crop row fitting. When the angle between the line connecting the crop and the previous crop center point and the vertical plane is greater than 30 degrees, the crop point is also taken as an interference point and is not involved in the crop row fitting.

In general, the DCGA-YOLOv8 model and threshold-DBSCAN proposed in this study can be well applied to multi-crop detection and crop row fitting in complex agricultural environments.

3. Present your research objectives more clearly: 1,2,3….

Responses:

This paper proposes a multi-crop navigation line extraction method based on improved YOLO-v8 and threshold-DBSCAN algorithm under complex agricultural environments. The detailed process and goals are as follows: (1) develop an improved YOLOv8 model DCGA-YOLOv8 combined with deformable convolution and global attention mechanism (GAM) to realize automatic and efficient detection of different crop rows in complex environments.; (2) apply a combined crop line fitting method based on DBSCAN cluster analysis, LSM and B-spline curve method to generate straight and curved crop lines; (3) construct a risk optimization function of the wheel model on the basis of the attitude of the agricultural machinery relative to the crop or the crop row. To make it more clear, we have added it in the revised manuscript.

4. Bring your conclusions in line with them.

Responses:

Aiming at the poor generalization and robustness of the existing navigation line extraction algorithm for multiple crops in complex farmland environment, this study proposes a new method based on improved YOLOV8 model, threshold DBSCAN clustering, least square method and B-spline curve was proposed to accurately detect multi-crops and crop rows. The specific implementations of the paper were as follows:

(1) a DCGA-YOLOv8 model was developed by introducing deformable convolution and GAM on the original YOLOv8 model for the detection of multiple crops. The F1-score and mAP value of the DCGA-YOLOv8 model for Cabbage, Kohlrabi and Rice were 96.4%, 97.1%, 95.9% and 98.9%, 99.2%, 99.1%, respectively. The comparative test results shown that the DCGA-YOLOv8 model is superior to the original YOLOv8, YOLOv5, Faster R-CNN and SSD methods in multi-crop detection.

(2) a distance threshold detection method was introduced to DBSCAN algorithm to exclude some image edge points and interfering points in terms of accurate clustering and grouping of crops in different rows. The correct clustering rate for Cabbage, Kohlrabi and Rice reached 98.9 %, 97.9 % and 100 %, respectively. LSM and cubic B-spline curve method successfully fit straight and curvilinear crop rows in the farmland environment.

(3) a risk optimization function of the wheel model was constructed to further improve the safety of the operation of the agricultural machinery between the crop rows.

We have added it in the revised manuscript.

5. I also asked to present more specifically and in detail the practical meaning of the study. What will be the impact of retrieving agricultural machinery navigation lines? To reduce crop losses, to reduce fuel consumption, to speed up harvesting? What is the economic effect (in quantitative units), due to what? This is very important, since the article will be published in the journal “Agriculture”.

Responses:

Just as you said, crop row detection has a significant impact on precise farmland management and intelligent agricultural decision-making. Accurate crop row detection and agricultural machinery trajectory planning can reduce the fuel consumption and production cost of agricultural machinery, and improve the utilization rate and operation efficiency of farm machinery. In addition, the optimized track of agricultural machinery can reduce crop losses during agricultural operations and improve the sustainability of agricultural production. To make it more clear, we have added it in the Introduction in the revised manuscript.

Author Response File: Author Response.pdf