You Only Look Once: Deep Learning Closes Technology Gap in Tart Cherry Counting

When it comes to counting tart cherries and estimating yield, USU researchers are adopting YOLO: You Only Look Once.

YOLO is an object detection algorithm that uses cameras to capture an image of the collected tart cherries and count the yield. Yield estimation is important to generate yield maps and aid in orchard management. Knowing how much fruit each tree produces can affect measures taken to improve soil, crop health and irrigation management to reduce water use while increasing yield.

Utah Water Research Laboratory graduate student Anderson Safre is taking YOLO to the next level. He used two versions of the algorithm, YOLOv8 and YOLO11, combined with a tracking algorithm to count tart cherries as they were harvested. The number output was then compared to the actual weights of the harvested fruit from individual trees to see how accurate the deep learning framework is at estimating yield and detecting individual cherries on the harvester.

“The fact that tart cherry orchards are mechanically harvested opens up a big opportunity for new technology, especially due to the repetitive nature of the process, which is perfect for automation,” Safre said. “Tart cherries have been mechanically harvested since the 1960s. To the best of our knowledge, no systems have been proposed to estimate yield during harvest.”

Safre saw a technology gap in tart cherry yield monitoring, and he aimed to address it by attaching cameras straight to the conveyer belt and hooking up the algorithms.

YOLO is widely used for fruit detection but requires different adaptations for different fruits. Safre explained that detecting small objects is a specific challenge for the algorithm and adjustments must be made for crops like tart cherries.

Safre used YOLOv8 and YOLO11 and compared their performance. YOLO 11 is the latest version, as of publication of Safre’s paper, and is largely similar to YOLOv8 with key enhancements. Safre had to choose different versions of the algorithms, which vary in accuracy and speed. For this study, he compared nano and extra-large versions.

Safre and his team found that while the larger versions slightly improved accuracy, they require significant computation power, impacting training time and estimation speed, and the difference in accuracy wasn’t huge. The smaller versions still performed well and have the added benefit of being deployable on microcomputers like the Raspberry Pi, making real-time fruit counting possible. They recommended that orchard managers make their choices based on the best trade-off for their individual situation.

Their research noted sources of uncertainty in the study, like the fact that the relationship between fruit count and fruit weight isn’t always linear. Some trees have lots of small cherries, while others have fewer but larger ones. Camera quality and angle on the conveyer belt also significantly impacted ability to detect the individual tart cherries, as well as pile-ups on the conveyer belt that hide cherries behind each other, also known as occlusion.

“Using this computer vision approach, we were able to estimate yield with an error of about 10 kg, which is very promising, especially given the large number of fruits in the videos and the occlusion challenges,” Safre said.

Orchard managers make the ultimate decisions on which algorithms and cameras to use to best fit their needs, but Safre has laid the framework to close the technological gap in yield monitoring. Orchard managers can use this method to closely track how many cherries each tree produces, revealing yield-limiting factors and helping them plan for workers, storage space and shipping needs.

Read more about Safre’s agriculture work under USU professor Alfonso Torres in their paper here.

Additional details about the patented technology is available on the USU Technology Transfer Services website, here. For commercialization interest, contact Alan.Edwards@usu.edu.