Journal of Experimental Agriculture International

Background: Tomato ripeness is identified by colour changes, but manual harvesting is often inaccurate. AI-based vision systems, especially deep learning models like YOLO, improve detection and harvesting efficiency.

Aims: In automated tomato harvesting, accurate classification and localization of mature fruits are essential for efficient picking operations. The present study aimed to real-time detection framework using YOLOv26 deep learning for automated tomato ripeness classification to support robotic harvesting applications. The objective was to achieve high detection accuracy and consistent performance under varying illumination conditions.

Study Design: An experimental research design was adopted involving dataset preparation, model development, training, validation, and real-time field evaluation. Tomatoes (hybrid NS 5037) were categorized into maturity stages beginning to change color: Immature, Turning, and Mature.

Methodology: A total of 1,000 annotated images were prepared using circular bounding box (C-BB) annotation to accurately represent fruit geometry. Circular bounding (C-BB) was employed to accurately represent the fruit's geometry, offering greater precision than traditional rectangular bounding. A YOLO (Yield-Only-Low) detection model was trained using the Ultralytics framework on a GPU-enabled platform, scaling all images to 640 × 640 pixels. Model performance was evaluated using precision, recall, F1 score, and average accuracy (AP). Real-time tests were conducted under sunlight and shade using an accelerated inference system with cuDNN.

Results: The results show that the proposed framework achieves 96.91% precision, 98.37% recall, 95.83% F1-score, and 97.85% AP under sunlight, while maintaining 95.47% precision, 96.44% recall and 95.67% AP under low light conditions.

Conclusions: These results demonstrate the proposed system's ability to detect tomato ripeness reliably and accurately under varying light conditions. And it has high potential for precision agriculture and robotic harvesting.