5.1 Beyond linear regression
The simplest yield models are linear in features and reach surprisingly good accuracy on well-managed staple crops in stable climates. Modern production models add non-linearity through gradient-boosted trees, Gaussian-process priors, and recurrent neural networks operating on time-series of phenological observations. The standard does not prescribe a model family; it prescribes the wire format for inputs and outputs and requires that any conformant model publish its feature-importance attribution per prediction.
5.2 Ensemble strategies
Single-model predictions are vulnerable to model-specific failure modes. A typical production ensemble combines (a) a process-based crop-growth model (e.g. APSIM, DSSAT) which encodes biophysical constraints, (b) a satellite-NDVI-driven empirical model which captures observed canopy state, and (c) a gradient-boosted tree on historical yield records which captures management effects. The ensemble's prediction is the weighted average; the weights are learned per-region against historical hold-out data.
5.3 Handling missing or late-arriving features
Real fields produce incomplete inputs. A farmer may not have a soil moisture probe; a satellite pass may be cloud-obscured for two weeks; a weather station may be offline for a critical phenological window. The standard requires that conformant models declare (a) the minimum feature set below which the model refuses to predict, and (b) the imputation strategy for partially-missing inputs. The output envelope carries both the prediction and the count of features actually observed versus imputed, so consumers can de-weight predictions made on heavily-imputed inputs.
5.4 Concept drift and model decay
Climate-change-driven shifts in growing-season length and
precipitation timing make a five-year-old yield model unreliable
in the present. The standard requires retraining cadence
metadata on every model envelope: the date of last retrain and
the data window used. A model retrained more than 24 months ago
against pre-2024 data is flagged with a stale
warning in conformance reports.
5.5 Domain-specific augmentations
For tree crops (orchard fruit, nuts, perennial vines) the standard
adds an alternate-bearing carry-forward feature: this year's
heavy crop predicts next year's lighter crop with strong empirical
evidence. For row crops (corn, wheat, soy) the standard adds a
planting-density feature distinct from area. For greenhouse crops
the standard substitutes integrated photon flux for solar
irradiance. These are documented as feature-set extensions,
referenced by feature_set_id.