Multi-Stage Tomato Flower Detection Using Deep Learning
Jennifer Dinh (Webster University), Ayanna Miles (Webester University)
Pollinator decline poses a significant challenge to greenhouse agriculture, particularly for crops such as tomatoes that rely on vibration-based pollination. Although robotic pollination systems have been explored, their effectiveness is highly dependent on accurate flower detection and stage identification. This work presents a multi-stage tomato flower detection framework using two state-of-the-art object detection models: YOLOv8m and RT-DETR-L. Unlike prior work that focuses on binary classification (e.g., flower vs. bud), we evaluate a three-stage scheme — Bud, Anthesis, and Post-Anthesis — provided by a publicly available dataset. Multi-stage classification has the potential to inform robotic decision-making, such as determining which flowers are ready for pollination, which should be revisited, and which are not yet suitable. The models are trained on this dataset and evaluated on both structured and real-world images. Results showed that YOLOv8m achieves higher detection performance with mAP@0.5 = 92.1%, compared to the 89.2% achieved by RT-DETR-L on the structured dataset, while performance drops to 71.1% and 64.2%, respectively, on real-world images. This performance gap highlights the complexity of flower stage classification and the limitations of current annotation practices. Inconsistent stage labeling and missing flower instances reduce the reliability of stage information; consequently, while this framework provides a reproducible pipeline for further evaluation, it is not yet ready for fully autonomous robotic pollination.