A novel dataset and deep learning object detection benchmark for grapevine pest surveillance

Area of Science:

• Agricultural Entomology
• Computer Vision
• Machine Learning

Background:

• Flavescence dorée (FD) is a major threat to grapevines, causing significant yield loss and economic damage.
• Current monitoring relies on labor-intensive manual identification of leafhoppers (Scaphoideus titanus, Orientus ishidae) on sticky traps.
• An automated pest detection system is needed to improve monitoring efficiency and disease management.

Purpose of the Study:

• To develop and evaluate an automated pest detection system for Scaphoideus titanus and Orientus ishidae.
• To create a comprehensive annotated dataset for training deep learning models.
• To compare the performance of YOLOv8 and Faster R-CNN for pest detection.

Main Methods:

• Collected and annotated over 600 images of S. titanus and O. ishidae from yellow sticky traps.
• Utilized computer vision and deep learning techniques for object detection.
• Trained and compared YOLOv8 and Faster R-CNN models, incorporating image pre-processing and data augmentation.

Main Results:

• YOLOv8 achieved high accuracy with an mAP@0.5 of 92% and F1-score >90%.
• Faster R-CNN demonstrated strong performance with an mAP@0.5 of 86%.
• The study successfully addressed dataset limitations and class detection challenges.

Conclusions:

• The developed automated system shows high potential for accurate and efficient detection of key FD vectors.
• This technology can significantly enhance Flavescence dorée management strategies in vineyards.
• The annotated dataset and trained models provide a valuable resource for future research and application.