Fresh market tomato is one of the most valuable crops in the US. However, production relies heavily on manual labor, which can account for over 30% of the total per-acre cost, with a large portion attributed to harvesting. In the southeast US, most tomato plants are staked and tied, and fruit are hand-harvested multiple times as they mature, increasing labor costs and operational inefficiencies. Compact Growth Habit (CGH) tomato varieties have a shorter stature that does not need to be staked and allow for more labor-efficient harvesting options, providing a promising alternative to traditional production. A key breeding objective for CGH tomato is to develop lines with a more concentrated fruit set (CFS), defined as a higher proportion of fruits reaching maturity synchronously. This trait would enable once-over harvesting, substantially reducing labor inputs while improving operational efficiency. Furthermore, the successful implementation of once-over harvest strategies in CGH tomatoes may facilitate the adoption of mechanized harvesting systems, addressing labor shortages and enhancing scalability in fresh market tomato production. This study aims to develop a computer vision model to automate detecting and quantifying tomato fruits and flowers in CGH breeding trials. High-resolution RGB images of top-view canopies were collected from experimental plots during the spring and fall seasons of 2024, capturing phenotypic variability across diverse environmental conditions and growth stages. The dataset is undergoing preprocessing, annotation, and augmentation to enhance model robustness. A YOLO-based object detection model will be trained to classify and quantify flowers and fruits. Model performance will be assessed using standard evaluation metrics, including precision, recall, and F1-score. By accurately detecting and quantifying fruits and flowers across developmental stages, this system will enable breeders to analyze flowering progression and identify CGH tomato lines with improved CFS, supporting the selection of varieties optimized for once-over harvesting. Preliminary model training using 1,370 training images, 116 validation images, and 335 test images in roboflow using YOLOv11 yielded promising results, with a mAP@50 of 94.7%, precision of 85.1%, and recall of 91.0%, demonstrating the model's potential to support phenotyping for concentrated fruit set. Future research will focus on enhancing detection accuracy, expanding dataset diversity, and integrating multispectral imaging techniques to optimize model performance and applicability across different environments.
