ResiYou: AI-driven residue management for growers

Growers face a trade-off they have to make under time pressure. They need enough crop protection to control pests and disease effectively – and must not exceed legal maximum residue levels. On top of that come retail secondary standards, which are frequently stricter than EU requirements.

Every application is therefore a decision about when and how much to spray, with three requirements at once: EU regulation, buyer requirements, and the crop's actual protection needs. Dose too cautiously and you lose harvest. Apply too late or too much and you risk a batch that cannot be sold.

At its core is an extensive database of field trials and residue measurements going back to the 1970s – a depth of historical context few projects have available.

Three further sources were added: new data collected and processed by commercial growers; official residue reports from regulatory authorities; and information on the chemical properties of active substances, environmental factors and historical weather data, to capture environmental influence on residue levels.

Residue trajectories from incomplete sample series. The task was not to reproduce a measurement but to predict the trajectory of a residue profile over time – from sample series with gaps. Incomplete data makes any trajectory modelling considerably harder.

A low signal-to-noise ratio from real conditions. Data from commercial growers is not generated under trial conditions. Separating meaningful patterns from noise required correspondingly sophisticated analytical techniques.

Chemical and environmental interactions. Residue levels depend on an interplay of substance properties, weather and timing of application. That complexity markedly increases the difficulty of accurate prediction.

Uncertainty as a requirement, not a shortcoming. Because of the inherent variability of agricultural environments and strongly imbalanced data across several dimensions, it was essential to make predictions with a stated degree of uncertainty. A point forecast without confidence would be worthless for a compliance decision.

Scalable infrastructure on AWS. The data-intensive processing steps need compute that scales with demand – the cloud infrastructure provides it.

Gradient-boosted trees for prediction. For predicting residue levels we use gradient-boosted trees, among the strongest methods for structured data in both regression and classification tasks.

MLOps across the whole lifecycle. From data preparation through to model deployment, MLOps practices are built in. That is not optional in this application: residue models have to remain accurate and relevant over years while varieties, products and regulations change.

CI/CD for ongoing operation. Continuous integration and deployment pipelines ensure updates reach users reliably and without interrupting service.

ResiYou provides an interface into which growers enter their data and from which they receive precise residue profiles for the active substances used, at the point of harvest. Specifically, the platform enables:

• planning of optimal spray schedules
• real-time predictions of residue levels
• verification of compliance with EU regulation and retail standards
• decision support for contractual obligations towards buyers

That turns the grower's dilemma into a calculable decision: not a compromise based on experience, but a forecast with stated uncertainty.

"Soon after the Supper & Supper team begun to work on the project we started to see the results: our predictive models increased the accuracy tremendously and the feed-back of our customers turned to be extremely positive." – José Luis Robles Martín, New Venture Lead EMEA – Horticulture, Bayer

The pattern – trajectory forecasts from incomplete measurement series, stated uncertainty, compliance checking against several rule sets at once – applies wherever a threshold decision has to be made before the measurement exists. In pharmaceutical manufacturing that is prediction of specification conformity; in the food industry, residue analytics along the supply chain.

The platform is in production use and is being extended continuously. Current priorities: integrating the environmental assessment tool from our sustainability project with Bayer so that growers see residue compliance and environmental impact within the same decision; moving the uncertainty estimates from model intervals to conformal prediction, which gives guaranteed coverage per individual case; and further automating data quality checks, since a growing share of incoming data comes from practice rather than from trials.