Models Overview - AgriHub360

AgriHub360 runs a set of models on each field. Each has its own page: what it predicts, the published science behind it, how it’s applied, which external models it touches, how it’s validated, and what it doesn’t do. Most of the models are mechanistic or rule-based. They apply published relationships, such as thermal time or the FAO-56 evapotranspiration equations, to the weather and readings for a field. None of them measures soil nutrients, and none of them calls a language model.

Models

Crop Development

Model	Type	Resolution	Predicts
Growth Stage	Mechanistic	Field	The current Zadoks stage of a planting and the date of the next, from accumulated degree-days
Drilling Window	Rule	Field	Whether soil temperature has held above a crop’s threshold for long enough to drill

Field Conditions

Model	Type	Resolution	Predicts
Spray Window	Rule	Field	Hours when delta-T, wind and rain allow spraying, and a frost risk
Water Stress	Rule	Field	Reference evapotranspiration and a banded drought-stress index
Nitrogen Timing	Rule	Field	Whether to apply or hold nitrogen, from growth stage, soil temperature, soil moisture and the forecast rain

Disease & Pest Risk

Model	Type	Resolution	Predicts
BYDV Risk	Mechanistic	Field	When accumulated day-degrees since emergence pass the treatment threshold
Fusarium Head Blight Risk	Published logistic, with a learned model scored against it	Regional	The probability of an FHB epidemic in the region, from hours of warmth, humidity and rain around flowering
Foliar Disease Risk	Mechanistic	Field	A pressure level per disease for Septoria tritici, yellow rust and brown rust, from hours of leaf wetness at infective temperatures

Language & Vision Models

A second group reads what the field models write and turns it into text, or reads a photo. Scout Photo Diagnosis names a disease or pest from a leaf photo. The Field Health Review synthesises a field’s readings, forecast, NDVI and model outputs into a status and actions. EUDR Document Drafting writes the due-diligence statement from the farm’s records.

Model	Type	Reads	Produces
Scout Photo Diagnosis	Vision model	A leaf photo, location data stripped and downscaled	A species from a curated list, a confidence and what the model saw
Field Health Review	Language model	Readings, the weather outlook, the soil sample, the NDVI capture and the field models’ outputs	A status, headline, factors and actions, each action with a condition the server checks
EUDR Document Drafting	Language model	Parcels and dates from the record, your answers and the farm’s metadata	Sections with headings and body text, rendered to PDF

How the Models Run

The field models share one pipeline. Every model reads the same field inputs: the weather, the readings, the soil sample, the planting record and the growth stage. Each input carries the name of its source. A model whose required input is missing skips the field and writes nothing for it, rather than filling the gap with a zero or an estimate. Each model’s pure core writes its own row for each run and raises advisories and notifications, so the history of stages, indices and risk values for a field is kept and the field screens read the latest.

Rules Every Model Follows

Skip, don’t approximate. If a required input is missing for a field, the model emits nothing for that field.
State the resolution. A model trained on regional data is labelled regional and is never shown as a per-field result.
Report skill against a baseline. A model that doesn’t beat its baseline is shown as its baseline only.
Show the uncertainty. If a number is shown, its interval or confidence is shown with it.
Cite the source. Every threshold traces to AHDB, a peer-reviewed paper or validated field data, named on the model’s page.
Respect coverage floors. A model over patchy data under-reads. Below a stated floor, it declines to answer.

Validation

Models are validated by leaving out a whole season, a whole region, or both, and scoring the held-out predictions against a named baseline. Random cross-validation isn’t used: nearby fields in the same season are correlated, so a random split leaks. Each model page states how it’s validated and which baseline it’s scored against.

Inputs

Inputs are a gridded weather archive and forecast, an hourly weather series, elevation, your devices and your soil samples. Nutrient values come from soil samples, not from a sensor reading. Each model page lists its inputs, their source, and what the model does when one is missing.

​Models

​Crop Development

​Field Conditions

​Disease & Pest Risk

​Language & Vision Models

​How the Models Run

​Rules Every Model Follows

​Validation

​Inputs