AI horticulture fruit counting system maps hidden crops in 3D

University of Canterbury researchers say an AI horticulture fruit counting system can find and measure fruit that is hidden by leaves, a step they say could make crop forecasts more accurate. The work matters for growers because harvest plans depend on knowing how much fruit is coming, what size it is and where it is on each plant.

The project is led by computer science professor Richard Green at UC Vision, with researcher Dr. Richie Ellingham developing the AI and computer vision technology, according to the university. Green said the system builds complete 3D models of plants, including fruit and plant structures blocked from normal view by foliage.

What is the UC Vision fruit counting system?

UC describes the system as a combination of cameras, lighting and AI tools that scan vines and orchard trees, then digitally remove leaves from the model to show what sits behind them. The goal is to create a usable 3D representation of each plant so computers and, eventually, machines can identify fruit and its position.

Green said factory robots have long depended on exact computer-aided design models, while orchards and vineyards are harder because plants vary from one another. According to UC, the new technology is designed to deal with that variation by modeling individual trees, vines, branches and fruit rather than relying on a standard plant shape.

The university said the system can identify apples, cherries and grapes, show the branch or cane they are attached to, and calculate size, volume and surface area. UC said repeated scans can also show how much a specific piece of fruit has grown over time.

Why fruit counts matter for growers

Green said growers now depend heavily on workers who sample, count and measure fruit before harvest, and he said that process can be inaccurate by as much as 23%. In UC testing, the new system produced fruit counts within about 2% to 3% of the correct total, according to the university.

Better estimates could affect labor, storage, packaging and waste, Green said. He said large producers can lose money if they prepare for the wrong crop size or fruit-size distribution, leaving them with too many or too few workers and materials at harvest.

• UC said the vineyard version uses two rows of cameras to capture images through dense vine canopies.
• The orchard version uses a taller camera frame that can scan trees of about 3.5 meters, or 11.5 feet.
• Green credited contributions from UC researcher Dr. Oliver Batchelor on AI algorithms, Ellingham on mechatronics, and research engineer Matt Mattar on commercial-grade software.

Green said detailed plant models could later help autonomous machines carry out pruning, thinning, spraying and harvesting. In cherries, he said a machine could target only fruit that has reached the most valuable size, then return later for the rest of the crop.

UC said cherries that are a few millimeters larger can be worth twice as much, which makes selective harvesting valuable for crop returns and planning. The university said the technology grew out of roughly 15 years of research backed by more than $32 million in government investment in computer vision, artificial intelligence and agricultural robotics.

Ellingham said the team has gathered 3D modeling data from 20 commercial farms to test the system under real vineyard and orchard conditions. According to UC, the researchers are now working to make the technology reliable enough for possible commercialization through HoloCrop.

Ellingham said farms working with the team are already asking when they can use the sampling tools. UC said HoloCrop is being set up to serve the fruit production chain with precision horticulture tools aimed at better data, less food-production waste and future robotics.

This story draws on original reporting from Phys.org.