Photo courtesy of Source.ag
Source.ag, a Dutch software and data platform for greenhouse horticulture, announced updates to the accuracy of its AI-powered Harvest Forecast model for tomatoes.
The model introduces changes to how it is trained, updated and improved, leveraging AI to continuously advance accuracy through real-world cultivation data. The newest generation model has been available to a small group of growers for field testing, with a full rollout imminent.
"The result is a model that handles climate variability that no individual grower could encounter alone," the company said in a news release. "The results are remarkable."
At the three-week forecast horizon, mean forecast accuracy has increased by 33% compared with the previous-generation model. The share of cultivations with forecasts more than 20% off target has declined by 25%. And severe outliers — cultivations where the forecast had big misses — have been reduced by 50%.
Behind these results is a change to how the model learns and develops. This next-generation model combines horticulture science with an AI model that learns from the setup and challenges that are unique to every greenhouse, accounting for all the ecological and logistical complexities that influence the daily harvest.
Apart from increased accuracy, the most significant practical change for growers is a substantial reduction in manual input. Several data processes that previously required regular manual registrations have been automated, removing a recurring operational burden while improving forecast accuracy
Rollout is phased, with an initial group of greenhouses already live and a broader expansion planned in the coming weeks. Full deployment to all tomato customers is expected in the coming months. The company said this model marks the beginning of a broader AI transformation for Source's growers and sales teams.
"What makes this meaningful is how the model improves over time," said Rien Kamman, CEO and co-founder of Source.ag, which was founded in 2020 by Kamman and Ernst van Bruggen. "The AI model learns from every cultivation running through Source.ag, so the model's performance compounds as more growers use the platform. That's the direction we've been building toward: AI that gets better with scale, not just with manual effort.”
“The most tangible change for growers is how much less the model asks of them," added Sebastiaan Vermeulen, data scientist at Source.ag. “We've automated several of the data processes that used to require regular manual input, which removes real friction from the grower's week. At the same time, we made a number of changes to how the collected data is used in the new model, leading to substantial improvements in the prediction accuracy. We’ve now made it easier than ever before to hit harvest commitments, realize a better price and reduce waste.”