ASHS 2025 Conference

Blueberry (Vaccinium spp.) production has been expanding globally, driven by its unique flavor and numerous health benefits. With increasing demand, it is essential to implement techniques aimed at improving yield in commercial blueberry fields. The distribution of blueberries across the canopy profile is a key factor linked to several management aspects, including the efficiency of mechanical harvesting, pest control strategies, and overall fruit quality. Understanding fruit distribution within the canopy is essential to optimize machine calibration, apply precision agriculture techniques for pest management, and incorporate technologies that enhance fruit quality and maximize production by reducing losses caused by uneven vertical distribution. This study aimed to determine the vertical distribution of blueberry fruits using RGB images and the YOLOv8s model. Data was collected from a commercial Southern Highbush blueberry field in Homerville, GA. Using a cellphone camera, a total of 200 images were collected from 40 plant samples, with each plant photographed from the top to the base at a consistent angle. The images were annotated and classified according to the fruit’s position on the plant (upper, middle, or lower third). The YOLOv8s model was trained on the labeled images and evaluated using standard metrics, including precision, recall, and Intersection over Union (IoU). The dataset was divided into 70% for training, 15% for validation, and 15% for testing, over 200 training epochs. The final outputs consisted of annotated images, performance metrics, and a summary table showing the number of plants analyzed and fruit concentration by canopy zone. Using the YOLOv8s model, we successfully mapped the spatial distribution of blueberry fruits across different canopy zones (upper, middle, and lower thirds). The model achieved moderate to high accuracy metrics, demonstrating good performance in fruit detection and localization. These results indicate the model's potential for practical field applications, although performance could be further enhanced by expanding the dataset and incorporating additional training cycles. Future work aims to test the model across different blueberry varieties and integrate it into a user-friendly web platform. These findings highlight the feasibility of using deep learning tools to support data-driven management decisions in commercial blueberry production systems.