Enhancing Grapevine Node Detection to Support Pruning Automation: Leveraging State-of-the-Art YOLO Detection Models...

Area of Science:

• Agricultural Robotics
• Computer Vision
• Viticulture Technology

Background:

• Automated pruning requires precise identification of cutting points on grapevines.
• Node detection on grapevine canes is essential for determining optimal pruning locations.
• Current methods face challenges in robotic manipulation and environmental perception.

Purpose of the Study:

• To propose and evaluate a novel method for grapevine node detection using state-of-the-art YOLO models.
• To assess the robustness and performance of YOLOv7, YOLOv8, YOLOv9, and YOLOv10 in diverse vineyard conditions.
• To create a publicly available dataset for grapevine node detection.

Main Methods:

• Training and validation of YOLOv7, YOLOv8, YOLOv9, and YOLOv10 models on grapevine images.
• Utilizing a public dataset with artificial backgrounds and real-world vineyard images from Portuguese regions.
• Evaluating model performance based on accuracy (F1-Score) and inference speed.

Main Results:

• All tested YOLO models successfully detected nodes in various grapevine datasets.
• YOLOv7 demonstrated the optimal trade-off between accuracy (70%-86.5% F1-Score) and inference speed (approx. 89 ms).
• The study generated a new, publicly accessible dataset of Portuguese vineyard images for node detection.

Conclusions:

• The proposed YOLO-based node detection method is efficient for real-time applications.
• YOLOv7 is identified as the most suitable model for autonomous robotic pruning systems due to its performance balance.
• This research facilitates the development of automated robotic pruning systems in viticulture.