AI Harvest Forecasters Promise 90% Accuracy for Fruit Growers

FruitCast, a UK startup selling harvest forecasts to fruit growers, says its predictions land within 10% of actual picked volume one week out and within 17% three weeks out. "We guarantee less than 20% error," the company states. That precision claim sits at the center of a small but expanding market for AI tools that tell farmers when to harvest — and getting the date wrong carries real costs.

Book seasonal workers for the wrong week and you pay for labor you don't need, or miss the window when prices peak. High-value crops such as strawberries and blueberries fluctuate wildly in price, which makes timing a direct profit question rather than a scheduling nuisance.

Joel Carter has seen the problem firsthand. At Okanagan Specialty Fruits, which grows apples across more than 1,250 acres of orchards in Washington State for sliced apple portions sold to hotels and schools, extreme heat derailed the first day of harvest last year. "It was like 38C… it's not safe for people to work in that heat," Carter recalls. "We had to stop at 10 o'clock in the morning."

He says AI models that forecast ideal harvest dates, factoring in weather, would help solve exactly this kind of problem. "You need to know more than just when your fruit is going to be ripe. How long do you have to pick it?" he says. "That's where these models are really helpful."

Okanagan is already experimenting with tractor-mounted cameras from Canadian firm Vivid Machines. The cameras capture imagery of apple trees as the tractor passes, and AI identifies buds, flowers or fruit in the footage. "Right now, Vivid is telling us crop estimates and harvest dates," says Carter. He notes the system is good at detecting very tiny flower buds that are hard to see with the naked eye.

The technology has a hard limit, Carter adds: forecast accuracy depends heavily on the quality of historical data fed into the system. "This isn't something where an AI can scrape the internet and figure out what's the average [yield] for Granny Smith," he explains. "It's going to be bespoke to your farm."

Apples offer some slack — the harvest window for Granny Smiths runs three weeks, according to Carter. Berries allow no such margin. "If a strawberry crop is on, you have to harvest it – otherwise your entire crop gets diseased very, very quickly," says Raymond Martin, co-founder and chief operating officer of FruitCast.

His company provides crop predictions for strawberries, raspberries, blackberries, blueberries and tomatoes. "We're moving on to grapes next year," Martin adds. FruitCast's model accounts for weather and irrigation conditions, and ingests footage of ripening fruit captured by drones, a smartphone carried on foot through a field, or a camera mounted on a farm vehicle.

Martin argues the value lies in scale, not novelty. Experienced farmers generally know when their fruit will ripen — but not necessarily across an entire farm spanning many acres, or one mixing outdoor and indoor growing areas. "We do exactly what the farmers could do but we just do it on a scale that they can't," he says.

This year tested the model. Hot weather and severe drought across much of the UK stressed fruit plants, pushing them into thermal dormancy and slowing fruit production, Martin notes.

Growers are testing the tools but withholding judgment. Angus Soft Fruits has worked with FruitCast, and its operations director Neill Finlayson says AI ripeness forecasting is not yet "a finished solution." "This journey is still ongoing and, whilst significant progress has been made, the industry remains some way from achieving a fully integrated forecasting ecosystem," he says. Driscoll's, the California-headquartered fruit seller with a large UK operation, confirmed to BBC News that some of its independent UK growers have used FruitCast's technology.

Academic researchers are pushing the analysis deeper. Many farmers already use handheld brix meters, which measure sugar content by tracking how much light refracts after passing through a liquid; some use infrared, avoiding the need to cut into the fruit. Yasaman Ghasempour at Princeton University says millimetre waves — high-frequency radio waves — can penetrate deeper. "They basically respond very well to humidity, water [and sugar]," she says. She and her students built a millimetre-wave ripeness detector that could serve farmers or even shoppers hunting for perfectly ripe fruit. When her students tested it at a local market in New Jersey, "[staff] there got kind of scared that we were doing something shady," she laughs. Fine-grained ripeness readings of unripe fruit could feed into future forecasts.

Cost is not necessarily the barrier. Kevin Wang at the University of Florida has developed a crop-counting tool that harvests imagery from $100 (£74) drones. Jing Zhang at North Carolina State University works on a system to automatically count blueberries captured in smartphone images of bushes.

Adoption is the harder problem. "Adoption is very complicated," says Zhang. "The grower has to have confidence in the research and whether or not it works." Wang points to another friction point: some farmers may resist sharing commercially sensitive information about fertiliser and irrigation plans with third parties and AI models.

Ben Palone, senior director of automation and commercialisation at Western Growers, an association representing farmers in the western US, says harvest forecasts have potential as "an optimisation tool." But he expects growers to keep humans in the loop. "Growers like to have people in the mix to make some of those very critical decisions – especially when it comes to harvests," he says.

For vendors like FruitCast and Vivid Machines, the near-term task is proving that 90% forecast accuracy holds up across enough farms — and enough drought-stressed seasons — to earn that confidence.