Skip to content

Systems and records

Agriculture yield correction dataset: keeping the reason a yield changed

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

An agriculture yield correction dataset keeps every yield value as its own version and never overwrites the first. Each later version carries a reason code such as calibration, moisture basis, scale reconciliation or area, links to its evidence (scale tickets, calibration logs, moisture readings, boundaries), records who made and approved it, and points to the version it replaces, so the final reconciled yield can be walked back and replayed through lots and deliveries.

System export

Agriculture yield correction

This page covers agriculture yield correction dataset: keeping the reason a yield changed.Illustrative

Key takeaways

  • Never overwrite the first yield value; every change is a new version pointing at the one it replaces.
  • Use a short list of reason codes, each with its own evidence, instead of free-text notes.
  • Copy the EPCIS and PROV pattern: the correction is a new statement and the original stays queryable.
  • Prove outcomes with records made outside the corrector, such as scale tickets, not with a status flag.
  • Replay a sample from delivery back to field and evidence; cases that fail the replay stay out.

How should agricultural data preserve the reason for a reconciled yield change?

Never overwrite the first number. Keep every yield value as its own version, and give each later version a reason code, the evidence it rests on (scale tickets, settlement sheets, calibration logs, moisture tests, area changes), the person who approved it, the time it was made and a pointer to the version it replaces. The final reconciled yield is then the last link in a chain anyone can walk back.

The reason is what makes the record useful to AI. A model that sees only the final yield learns nothing about why combine estimates drift, how moisture and shrink move a number, or when a field boundary was wrong. A model that sees the chain can learn to flag a suspect estimate, propose the right reconciliation and explain it, which is the job a farm manager, agronomist or grain merchandiser actually does.

This matters well beyond model training. Crop insurance production reports, food traceability records and landlord settlements all depend on someone being able to show how a number was reached. A correction lineage is the same evidence, kept in a form a machine can read.

Why does a field's yield number change after harvest?

A yield changes after harvest for a short list of reasons, and each one has its own evidence. Naming the reason as a code, not free text, is the single step that turns a pile of edits into a dataset.

Standards already separate several of these. The AgGateway ADAPT data model groups harvest yield representations that distinguish wet and dry mass and volume, stores calibration factors against a specific meter and time scope, and records loads with a load number, a load type such as tank, truck or bale, a quantity and destinations. Those are the hooks a reconciliation needs.

Why does a field's yield number change after harvest?
Reason codeWhat happenedEvidence to linkWhat must be kept
CALYield monitor recalibrated after harvest startedCalibration log with meter and time windowPre- and post-calibration values for the affected window
MOISTWet weight converted to a dry or standard-moisture basisMoisture readings and the conversion basis usedWet value, moisture and the basis, not only the dry result
SCALEMonitor total reconciled to delivered or stored weightScale tickets, settlement sheets or bin measurementsMonitor total, ticket total and the allocation method
AREAHarvested area correctedRevised boundary or as-harvested coverageOld and new area with the source of each
SPLITLoads from several fields shared a truck or binLoad records and the split rule appliedField shares and who set them
ENTRYKeying or unit error fixedOriginal entry and the corrected entryBoth values and the corrector
VOIDA recorded load or pass did not occurNote explaining why the record is voidThe void record itself, never deleted

What do traceability and provenance standards say about correcting without erasing?

Both say the same thing: record the correction as a new statement that points at the old one, and leave the old one intact. That principle is what a yield correction dataset should copy.

In GS1's EPCIS standard, an event carrying an error declaration asserts that a prior event's claims are in error as of a declaration time; the prior event is not modified, and later queries return both the prior event and the error declaration. The Core Business Vocabulary gives two error reasons: did not occur, used when there are no corrective events, and incorrect data, where later events can describe what actually happened and are linked through corrective event IDs.

The W3C PROV ontology offers the general form: wasRevisionOf links a revised entity to the original it revises, as a particular case of derivation. A yield version that wasRevisionOf the earlier version, with an attached reason and evidence, is exactly that pattern.

Mapped to farm records, a voided load is the did-not-occur case and a reconciled weight is the incorrect-data case with a corrective record. The mapping keeps the vocabulary portable across buyers who already work with traceability data.

Which fields does every yield version need to carry?

Every yield version is one row with the fields below. The structure is deliberately small so a farm, cooperative, grain handler or agronomy service can fill it from records it already keeps.

At a glance

Which fields does every yield version need to carry

Three AI tasks follow directly from rows built this way. Given version 1 and the season's logs, flag whether the estimate is likely to need correction and why. Given the evidence, propose the next version with its reason code and the records it rests on. Given a finished chain, explain in plain language how the accepted number was reached. Each task has an answer key in the chain itself, which is what makes the records worth more than a final yield table.

Which fields does every yield version need to carry?
FieldMeaningRule
record_id and version_noStable ID for the field-season-crop, plus a sequential versionVersion numbers never reused
field_token, season, cropWhich field, year and crop the value belongs toField token maps to the farm map, not to an owner name
value and unit_basisThe yield value and whether it is wet, dry or standard moistureUnit basis required on every version
revises_versionThe version this one replacesEmpty only on version 1
reason_codeOne code from the agreed listOne code per version; combine changes into separate versions
evidence_refsIDs of scale tickets, calibration logs, moisture tests or boundary filesAt least one for every reason except ENTRY
made_at, made_by_role, approved_by_roleWhen and by whom the change was made and approvedapproved_by differs from made_by for SCALE and AREA
lot_or_bin_linksLots, bins or traceability lot codes the yield fedRequired when the crop left the farm
statusdraft, accepted or supersededExactly one accepted version per record at any time

Worked case (illustrative): a combine estimate reconciled to scale tickets

Every figure here is illustrative; no real farm or business is described. The task starts from a single corn field whose combine reported its harvest on a wet basis on the day of harvest.

The linked evidence arrives over several weeks: the operator's calibration log the same evening, moisture readings with each load, and the elevator's scale tickets once the grain is delivered. The decision at each step is a new version, approved by the farm manager, and the verified outcome is the match between the accepted version and the elevator's totals, not the fact that someone marked it accepted.

Note what the lineage does not contain: a single edit that rolls calibration, moisture and scale reconciliation into one number. Splitting them is what lets a model learn which kind of error a given pattern points to.

Worked case (illustrative): a combine estimate reconciled to scale tickets
VersionChange from previousReason and evidenceStatus
v1Starting value from the yield monitor, wet basisHarvest log for the combine and daySuperseded
v2Down 3 percent for the last 2 hours of harvestCAL: calibration log shows the factor updated mid-afternoonSuperseded
v3Converted to standard-moisture basisMOIST: moisture readings from the loads and the basis usedSuperseded
v4Down a further 2 percentSCALE: 14 scale tickets from the elevator, totals matched to loadsAccepted
v5 (rejected)Proposed up 6 percent to match last seasonNo evidence; reason given as expected yieldDraft, never accepted

How do you walk a delivered load back to the yield version it used?

It can, if each version names the loads and lots it fed. Starting from a lot or shipment, a reviewer should be able to walk back to the field, the season, the accepted yield version and the evidence behind it, and forward again to the delivery records.

At a glance

How do you walk a delivered load back to the yield version it used

The FDA's Food Traceability Rule supplies the shape for foods on its Food Traceability List. It defines critical tracking events, including harvesting, cooling, initial packing, shipping, receiving and transformation, with key data elements for each; a traceability lot code is assigned at initial packing or transformation, and growers keep a farm map naming each field and its location. A yield record that carries the same field names and lot codes can be joined to those events without guesswork.

Grain moved in bulk rarely carries lot codes, so the join usually runs through load numbers, bins and scale tickets instead. The replay test is the same: pick any delivery, recover the field share and the yield version used, and check that the evidence still supports it.

  • Replay step 1: choose a delivery or lot at random from the season.
  • Replay step 2: follow its load or lot link back to the field record and accepted version.
  • Replay step 3: open each evidence reference on that version and confirm it supports the change.
  • Replay step 4: confirm no superseded version is still feeding a downstream total.
  • Record the result per case; a failed replay excludes the case from training.

Which record proves a yield version, and which only claims it?

The outcome is proven by a record created outside the person making the correction, not by the status field that says accepted. An accepted flag only shows someone clicked a button; a scale ticket or settlement sheet shows grain crossed a scale.

Federal crop insurance makes the same distinction. The Common Crop Insurance Policy Basic Provisions define production records as acceptable verifiable records or acceptable farm management records, supporting the production reported. The dataset should label which kind of record proves each version, because that is the difference a model needs to learn.

Which record proves a yield version, and which only claims it?
What you want to proveRecord that proves itNot proof on its own
Grain quantity deliveredScale tickets or settlement sheets from the receiverMonitor total or a spreadsheet sum
Calibration was appliedCalibration log tied to the meter and time windowA note that the combine was calibrated
Moisture basis is rightMoisture readings and the stated conversion basisA dry yield with no wet value
Area is rightRevised boundary or as-harvested coverage fileA rounded acreage typed into a form
The decision was approvedApproval record with role and timeStatus field set to accepted

Which corrections are unfit to become training examples?

Corrections with no evidence, or ones made to hit a target, are the counterexample. Version 5 in the worked case is one: a proposal to raise yield to match an expected figure, with nothing behind it. Keep such proposals as labelled negative examples if the evidence of rejection is clear; otherwise leave them out.

Hold a case back, or ask for specialist review, when:

  • The original value was overwritten and only the final number survives.
  • Loads from several fields were commingled and nobody recorded the split rule.
  • The record is part of an open insurance claim, appraisal or settlement dispute.
  • Landlord, tenant or crop-share arrangements put the yield data under another party's control or consent.
  • Field tokens could be re-linked to an identifiable farm through location or boundary files.
  • A food safety or quality hold, rather than a measurement issue, drove the change; that needs the food safety lead's review.

Can a farm business license its yield correction histories?

Farms, cooperatives, grain handlers and agronomy services that kept their versions and evidence hold a rarer asset than yield maps alone: the reasoning behind each number. Farmer consent, landlord rights and data agreements with equipment and agronomy providers decide what can be licensed, and counsel should confirm them first.

Once those consents are clear, a grain business can take its correction chains to SourceX, which handles the rights review, contract, preparation, delivery and payment for records licensed to AI developers. A short assessment, with no files sent, tells a farm or cooperative whether it qualifies; nothing ships before it approves the scope and signs, and a multi-season archive can stay in its own storage or go out on encrypted drives.

Frequently asked questions

What is an agriculture yield correction dataset?

It is a set of field-season yield records that keeps every version of the yield number, the reason each change was made, the evidence behind it and who approved it. It lets AI learn how and why yields are reconciled, not only what the final yield was.

Why not just keep the final reconciled yield?

The final number hides calibration drift, moisture conversion, scale reconciliation and area fixes. Without the earlier versions and reasons, nobody can check the result or teach a model to spot a suspect estimate.

What evidence supports a yield correction?

Scale tickets or settlement sheets for delivered quantity, calibration logs for monitor adjustments, moisture readings for basis changes and revised boundaries for area changes. Each version should reference at least one such record unless it fixes a keying error.

How does EPCIS handle corrected records?

An EPCIS error declaration states that a prior event was in error as of a declaration time without modifying it, and both remain queryable. Corrective events that describe what actually happened are linked by their event IDs.

Can yield data be shared without identifying the farm?

Sometimes. Replace field and farm names with tokens and be careful with boundaries and coordinates, which can re-identify a field. Consent and contract terms with landlords and service providers must be checked first.

Sources

  • In EPCIS, an event with an error declaration asserts a prior event's claims are in error as of the declaration time; the prior event is not modified and queries return both. Source
  • The GS1 Core Business Vocabulary defines the error reasons did not occur and incorrect data, the latter linked to corrective events. Source
  • The AgGateway ADAPT model records loads with a load number, load type (such as tank, truck or bale), load quantity and destinations. Source
  • The ADAPT model stores calibration factors against a specific meter and time scope. Source
  • ADAPT's harvest yield representation group distinguishes wet and dry yield mass and volume. Source
  • W3C PROV-O defines wasRevisionOf, a derivation whose result is a revised version of an original. Source
  • The FDA Food Traceability Rule sets critical tracking events and key data elements, traceability lot codes and a farm map naming each field. Source
  • The Common Crop Insurance Policy Basic Provisions require production records to be acceptable verifiable records or acceptable farm management records. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies