top of page

Wageningen Drone Counts Orchard Fruit During a Single Flight

Writer: AgriLinkage Production
AgriLinkage Production
2 hours ago
3 min read

A Wageningen University doctoral project has moved orchard drone monitoring closer to producing useful fruit counts during the same flight, rather than hours or days later. The work delivered apple counts for individual trees within minutes and kept onboard processing fast enough to match the arrival of new images.


That matters because thinning, crop protection, labour planning and harvest decisions all depend on knowing how much fruit trees are carrying. A delayed map may describe the orchard accurately but arrive after the practical decision has already been made.


Graphic showing quantified results from Wageningen orchard drone research
Original Agrilinkage graphic based on the Wageningen doctoral thesis.


What the research achieved


The thesis by Zhen Cao, awarded on 16 November 2026, brought together several methods intended to reduce the computing burden that has kept much aerial crop analysis off the drone and outside the flight window. In a vineyard pipeline, a lightweight detector selected useful frames and sent only compressed regions of interest to an edge server. More than half of the incoming frames were discarded as redundant while the system kept pace with normal vineyard flight speeds.


A separate test examined fully onboard detection of very small grape bunches. Raising image resolution helped until a practical knee near 1920 pixels. A Transformer detector found tiny bunches 26 percent more accurately than a convolutional model, but it ran four times slower. The result exposes the central engineering tradeoff: a more accurate model can still be less useful in the field if it cannot finish before the next images arrive.


For apples, the system flew along two opposite sides of a commercial orchard row, detected trees and fruit onboard, located each tree by direct geographic positioning and combined both views. It avoided the slower photogrammetric reconstruction and tracking steps common in post flight workflows. The two sided count reached 82 percent of the upper limit set by the fruit detector itself.


The full methods and results are documented in the Wageningen University research record and the linked doctoral thesis.


The biggest limit is still visibility


The numbers do not mean a drone can now deliver a perfect harvest forecast. Apples hidden by leaves remain invisible from the air, and very small fruit can occupy too few pixels for reliable detection. The 82 percent result is measured against a detector based upper bound, not against a complete manual count of every apple on every tree.


This distinction is commercially important. A grower deciding whether to deploy the system needs to know whether the count is stable enough for a specific task. Relative differences between orchard blocks may be useful for scouting or labour allocation even when the absolute total is incomplete. Contracting, storage planning and final yield forecasts demand a higher and independently validated level of accuracy.


What would make the system ready for growers


The next evidence should come from repeated commercial flights across cultivars, canopy structures, light conditions and crop stages. Researchers and suppliers also need to report energy use, processing latency, failed detections and the cost of the onboard computer. Cao reviewed 77 studies and found that latency and energy were rarely reported even though both determine whether a small drone can sustain the workload.


A practical product would also need a simple path from flight to decision. Per tree counts should flow into orchard maps, work orders and harvest plans without requiring a grower to rebuild the analysis. The value is not the drone image itself. It is a timely estimate that changes a farm decision.


What to watch next


The decisive tests will be whether two sided flight plans remain fast on larger orchards, whether estimates hold after pruning or weather changes, and whether better cameras or models close the visibility gap without making the aircraft too heavy or the processing too slow. The thesis shows that same mission output is technically plausible. It does not yet establish a universal commercial counting standard.


Cover image: original editorial illustration created for Agrilinkage. It does not depict the Wageningen trial.

Comments


bottom of page