University of Canterbury team reports 3D crop scans that model fruit hidden by foliage

Crop estimates influence operational choices made before harvest, including how many workers to schedule and how much packaging and storage capacity to arrange. UC Vision says manual sampling, counting and measuring can produce inaccuracies of up to 23%; a more accurate count could therefore reduce mismatches between expected and actual crop volumes. Plant-by-plant measurements may also help growers sort planning around fruit size, which the source notes can affect cherry value. These are prospective operational benefits, however, and not evidence yet of savings or waste reductions achieved in commercial use.

The work addresses a persistent constraint in horticulture automation: fruit and plant structures are often obscured by dense leaves, while individual trees and vines vary rather than conforming to a fixed factory layout. UC Vision combines cameras, lighting and multiple AI systems to remove foliage digitally and reconstruct plant models. Its mobile systems scan along orchard and vineyard rows; the orchard configuration can scan trees about 3.5 meters tall, while the vineyard setup uses two rows of cameras. The project builds on about 15 years of University of Canterbury research backed by more than $32 million in government investment in computer vision, AI and agricultural robotics.

The reported accuracy is limited to fruit-count testing: the source says counts were within about 2% to 3% of the correct figure, but does not provide the number of scans, crop-specific results, comparison method or error ranges for size and volume measurements. It also does not document commercial deployment, costs, performance across seasons, or whether the system maintains accuracy in every orchard and vineyard condition. Reliability remains an unresolved next step as the team prepares for potential commercialization.