Why Sensor Provenance Matters in Agricultural AI

Sensor provenance is the evidence trail that explains where an observation came from, when and where it was collected, how it was calibrated or transformed, and what uncertainty remains. Agricultural AI needs that trail because a plausible value can still describe the wrong place, the wrong moment, or a processing step the user cannot verify.

Agriculture produces heterogeneous evidence. A greenhouse may contain environmental sensors, controller histories, crop-work records, images, point clouds, laboratory results, and operator notes. Each source has a different clock, spatial resolution, calibration history, and failure mode.

Putting those values in one database does not make them one coherent observation.

A value without context is not enough#

Consider a temperature reading. The number alone does not answer:

• which sensor recorded it;
• where the sensor was located;
• which timezone or clock produced the timestamp;
• whether the device was calibrated;
• whether the value was raw, averaged, converted, or interpolated; or
• whether it applied to the crop event being investigated.

The same problem becomes more consequential with spatial evidence. A point may be valid in its source frame but misleading when placed in a greenhouse scene using an unsupported transform.

Provenance keeps those questions attached to the value.

The minimum provenance record#

For Streamline’s current reconstruction work, a useful evidence record includes:

1. Source identity. The system, file, device, record, or capture that produced the observation.
2. Recorded time. When the source says the observation occurred.
3. Ingestion time. When the reconstruction received it.
4. Physical context. Facility, zone, bed, crop area, or equipment association.
5. Units and representation. The source unit and any normalized form.
6. Transformation history. Conversions, coordinate transforms, filters, or derived values.
7. Quality and uncertainty. Missing calibration, clock ambiguity, stale data, or conflicting evidence.
8. Integrity information. Enough information to detect whether the source changed unexpectedly.

The exact schema can vary. The principle does not: the displayed or modeled value must remain traceable to the evidence that produced it.

Provenance before fusion#

Sensor fusion is often described as a way to create a more complete state estimate. That can be useful only when the inputs are compatible.

Before combining observations, a system should ask:

• Were they captured close enough in time for the question being asked?
• Do the coordinate frames and calibrations support registration?
• Are the relevant parts of the greenhouse unchanged between captures?
• Does one source have a known quality problem?
• Can the fused result retain the identity of its inputs?

In the GH1 / Bed 4 evidence, the reconstruction exposed an approximately 43-minute mismatch between spatial observations. The available synchronization and calibration evidence did not support presenting them as one moment, so Streamline refused the fusion.

That is provenance affecting system behavior, not provenance stored as decorative metadata.

Provenance before machine learning#

An agricultural model can learn a stable relationship that is operationally false if the training data repeatedly mixes locations, crop stages, treatments, or time windows.

For example, a pattern associated with genotype may actually reflect environment, crop work, or harvest timing. Regulated cannabis cultivation sharpened Streamline’s focus on this problem: cultivar identity sets biological potential, but observed expression remains entangled with conditions and operations.

Preserving those variables does not prove causation or guarantee repeatable genetic outcomes. It creates the evidence structure needed to test whether a relationship survives across cycles and contexts. The proposed World Biological Outcome Model extends that question beyond the current reconstruction.

Provenance before action#

The consequences increase when software moves from description to recommendation or control.

A reviewable recommendation should identify:

• the observations it used;
• the age and quality of those observations;
• the rule or model that produced the recommendation;
• the expected result and measurement window;
• the operator who approved or rejected it; and
• the evidence used to verify what happened afterward.

An automated action needs additional safety, authority, and completion checks. The operator-in-the-loop safety framework describes those later requirements. They are not capabilities of the current Streamline OS implementation.

What provenance does not solve#

Provenance can show what happened to data. It cannot make an inaccurate sensor accurate, recover a missing calibration, prove a causal biological relationship, or guarantee that an action is safe.

It can, however, stop those limitations from disappearing inside a polished interface.

That makes provenance a foundation for agricultural intelligence rather than another farm dashboard. Intelligence requires context that can be challenged, not only values that can be displayed.

Related Streamline field notes#

• Reconstructing a Greenhouse Event from Telemetry and LiDAR
• How Streamline Detected a 43-Minute Capture Gap
• What Is a Greenhouse Digital Twin?
• Read-Only Digital Twins Before Greenhouse Automation

Sources and scope#

This field note describes Streamline’s evidence-design principles and the behavior demonstrated by the public GH1 / Bed 4 case study. It does not claim that the current implementation performs predictive agronomy, causal inference, live recommendations, or equipment control.