Field data verification: what makes a field image usable

What is field data verification?

It is the gate every field asset passes before analysis. Verification establishes that the image is what it claims to be, that it may be used for this purpose, and that it is technically fit for the question being asked of it.

Verification runs before inference rather than after, because a model cannot tell you that its input was unusable. It will return a plausible answer on a blurred frame from an unregistered rig with no rights record, and nothing downstream will flag it.

What metadata must a field capture carry?

Enough to place the image in a specific field, on a specific crop, at a specific time, taken by known equipment in a known configuration. Metadata added afterwards from memory is not provenance.

Precise field coordinates and other sensitive operational details are handled under a signed agreement and an approved transfer process. They are never submitted through the public forms on this site.

What makes an image invalid for a decision?

Any failure of rights, registration, calibration, or measured quality. Invalid is a binary state, not a low score: the asset is excluded from the decision rather than weighted down within it.

• Rights do not cover the intended use, or cover analysis but not training.
• The asset does not belong to a registered capture session.
• Camera calibration is unknown, stale, or inconsistent with the recorded rig.
• Motion blur at working speed leaves the target unresolvable.
• Exposure fails: blown highlights in sun or crushed shadow inside the canopy.
• Occlusion by leaves, dust, water on the lens, or crop residue over the target area.
• Ground resolution at the target is too coarse for the life stage in question.
• Coverage does not span the area a decision would be asked to cover.

Treating invalid assets as merely weak is the failure mode worth naming, because it is the convenient one. Down-weighting a bad image still lets it influence an outcome, and still leaves a record implying the decision rested on adequate evidence.

Why are data rights part of verification?

Because a rights question that is not answered before use gets answered after, by someone else. Training eligibility is intended to be an enforceable state in the platform, not an understanding held between two people.

• Analysis rights and training rights are distinct. A partner may permit one and withhold the other.
• Rights have scope and duration, and both are recorded rather than assumed to be permanent.
• A dataset built from assets whose rights are unclear cannot be relied on later, whatever its quality.
• Where rights are withdrawn, lineage is what makes the affected models and datasets identifiable.

A field program begins only under a separate signed agreement and an approved data-transfer process. Verification enforces what that agreement says; it does not replace it.

What is field-image provenance?

The verifiable origin and handling history of an image: where and when it was captured, by what equipment, under what agreement, and what has happened to it since. Provenance is established at capture, not reconstructed later.

That timing constraint is the whole point. Once a folder of images has been copied between machines, renamed, and re-encoded, no amount of later diligence recovers the field, the rig, or the agreement it came from. Assets are ingested immutably and referenced thereafter, so that a decision points at the original rather than at a copy of a copy.

How does verification affect training data?

It determines what may enter a dataset at all. Rights, registration, calibration, and quality are checked before an asset becomes training data, and the resulting dataset lineage records which assets and adjudications composed it.

Unverified training data is the most expensive kind of shortcut available here, because its consequences surface much later and cannot be localised. A model trained on assets of unknown provenance cannot be defended, cannot be corrected precisely, and cannot be cleanly separated from the assets whose rights turn out not to cover training.

What does verification not check?

It establishes that an asset is admissible, not that the biology in it has been read correctly. A perfectly verified image can still be labeled wrongly or analyzed by a model outside its evaluated envelope.

• It does not confirm species or life-stage identity — that is detection and adjudication.
• It does not establish field-scale pest pressure, which depends on sampling design.
• It does not certify that a decision produced from the asset is correct.
• It does not substitute for the grower's own records or for any regulatory requirement.