Deep learning based approach for actinidia flower detection and gender assessment

Area of Science:

• Agricultural Robotics
• Computer Vision
• Plant Science

Background:

• Pollination is vital for subsistence crops like Actinidia (kiwifruit), which presents unique challenges due to dioecious flowering and lack of nectar.
• Attracting natural pollinators is difficult, driving interest in artificial intelligence (AI) and robotic solutions for pollination, especially in suboptimal conditions.

Purpose of the Study:

• To address the lack of gender-based flower detection and underutilization of deep learning in Actinidia pollination.
• To evaluate the performance of deep learning models (YOLOv8, YOLOv5, RT-DETR, DETR) for accurate gender-specific detection of Actinidia flowers.

Main Methods:

• Developed a manually annotated dataset of Actinidia flowers categorized by gender.
• Evaluated four pretrained deep learning models using k-fold cross-validation for flower detection and gender determination.
• Assessed model performance based on precision, recall, F1 score, and mean Average Precision (mAP).

Main Results:

• The DETR model demonstrated the most balanced performance, achieving high precision (89%), recall (97%), F1 score (93%), and mAP (94%).
• The study highlights the effectiveness of deep learning in accurately identifying and differentiating genders of Actinidia flowers.
• DETR's robustness was noted in handling complex detection tasks under varied environmental conditions.

Conclusions:

• Deep learning models, particularly DETR, show significant potential for enabling precise, gender-specific robotic pollination in Actinidia.
• This research paves the way for developing advanced AI-driven robotic pollination systems to improve crop yield and efficiency.