Why process models rather than machine learning?

Process models can be interrogated: a reviewer can ask which physiological mechanism drove a result. Machine learning is used where it improves inputs, not as the load-bearing yield mechanism.

Trust

Methodology

Phoeniks uses process-based crop models rather than a single statistical fit. APSIM, DSSAT and ORYZA each simulate crop growth from soil, weather and management inputs; running them as an ensemble exposes model disagreement, which is reported as part of the confidence band instead of being averaged away.

What this document commits to

Model family: Process-based ensemble
Models run: APSIM, DSSAT, ORYZA
Pipeline steps: Six, from ingest to ledger
Uncertainty treatment: Model disagreement reported, not averaged

Applied by

Key terms on this page

Pipeline, end to end#


1. Ingest	Satellite, weather and registry sources are pulled on schedule and versioned
2. Locate	Parcel geocoding and boundary resolution, with environmental overlays applied
3. Constrain	Crop catalogue and can-grow matrix narrow the plausible crop set
4. Simulate	The process-model ensemble runs over the property's history
5. Score	The scoring engine produces a comparable score and confidence band
6. Record	Model version, input snapshot and scenario are written to the ledger

Stated assumptions#

Management practice is inferred, not observed. Where it materially drives the result, this is flagged.
Weather reanalysis carries its own uncertainty, which propagates into the confidence band.
A modelled potential is not a forecast of an individual season's realised yield.

Frequently asked

Why process models rather than machine learning?: Process models can be interrogated: a reviewer can ask which physiological mechanism drove a result. Machine learning is used where it improves inputs, not as the load-bearing yield mechanism.

Trust