Deep learning framework for fruit counting and yield mapping in tart cherry using YOLOv8 and YOLO11

Abstract

Object detection for fruit counting has significant potential for orchard yield estimation. Tart cherries are mechanically harvested, creating opportunities for developing new yield mapping technologies. However, there is a lack of dedicated technologies for this purpose, motivating the evaluation of computer vision-based approaches in tart cherries. In this study, we compared the nano (n) and extra-large (x) configurations of YOLOv8 and YOLO11 for tart cherry detection and fruit counting on the harvester. The models demonstrated robust performance, even in high object density conditions, with YOLOv11x achieving a mAP50 of 0.92. While YOLOv8n and YOLO11n produced similar detection results, YOLOv8n had a faster inference time, making it more suitable for real-time applications. Fruit counting was performed using a combination of YOLO models and the BoT-SORT tracking algorithm. The resulting number of fruits was compared to the actual weights of harvested fruit from individual trees. The results indicated a linear relationship, with YOLO11x achieving an R² of 0.62 and an RMSE of 10 kg. To the best of our knowledge, this is the first study to evaluate object detection and fruit counting performance in tart cherries during harvest. Additionally, we introduce a new dataset with annotated cherries on the conveyor belt of the harvester which can support further research and development. This approach addresses the existing technology gap in yield monitoring for tart cherry orchards, facilitating the application of precision agriculture and site-specific management strategies in the industry.