Reconstructing a Greenhouse Event from Telemetry and LiDAR

A defensible greenhouse reconstruction does not begin by merging every available file. It begins by identifying each observation, locating it in the facility, placing it on a shared timeline, retaining its calibration and transformation history, and refusing fusion when the evidence does not support one synchronized state.

Environmental telemetry and LiDAR answer different questions.

Telemetry can describe temperature, humidity, equipment state, or another measured variable at recorded times. LiDAR-derived spatial evidence can describe geometry at the moment of capture. A greenhouse event becomes inspectable only when those observations can be compared without pretending they were collected under identical conditions.

Start with the event boundary#

Before reconstruction, define the narrow event being investigated:

• the facility and crop location;
• the start and end of the recorded interval;
• the operator question;
• the available data sources;
• the clocks, units, and coordinate systems involved; and
• the conclusion the current evidence is allowed to support.

The Streamline OS platform currently applies this approach to recorded evidence. It is not a live data pipeline or an equipment-control system.

Step 1: identify every input#

Each input should carry more than a value. The reconstruction needs a source reference, capture or record time, applicable location, unit, quality information, and transformation history.

For a spatial observation, that also includes the coordinate frame and the calibration or registration evidence needed to compare it with another capture. If those details are unavailable, the absence is part of the result.

This is the role of sensor provenance in agricultural AI: keep enough history to decide whether two observations are compatible before a model learns from them or an operator acts on them.

Step 2: normalize without erasing the original#

Recorded values often need unit conversion, timestamp normalization, or coordinate transformation. A reconstruction should preserve both the source value and the derived representation.

That separation allows a reviewer to ask:

• Was this temperature recorded directly or converted?
• Which clock produced the timestamp?
• Which transform positioned this point in the scene?
• Did the source file change after ingestion?
• Is this element observed, derived, simulated, or unavailable?

Normalization makes observations comparable. Lineage keeps the comparison auditable.

Step 3: build a timeline before building a scene#

A three-dimensional scene can make asynchronous observations look simultaneous. The timeline has to come first.

In Streamline’s GH1 / Bed 4 reconstruction, the inputs included 51 recorded greenhouse telemetry records and approximately 180,000 points in the reconstructed spatial scene. When the spatial capture times were placed on the same timeline, the workflow exposed a mismatch of approximately 43 minutes.

That gap was not a rendering problem. It was evidence that the captures represented different moments.

Step 4: test whether fusion is supported#

Combining spatial observations requires more than proximity. The reconstruction needs adequate evidence about time alignment, calibration, and coordinate relationships.

The GH1 / Bed 4 slice did not have sufficient support to treat the mismatched observations as one calibrated instant. Streamline therefore kept them distinct. The case is explained in How Streamline Detected a 43-Minute Capture Gap.

Refusal is a valid system outcome. It prevents a clean-looking visualization from silently becoming a false measurement.

Step 5: render state with provenance attached#

Once observations have been placed and evaluated, the scene can show:

• the facility and bed context;
• the recorded telemetry interval;
• the spatial observation associated with its capture time;
• derived transformations and their lineage;
• unavailable or incompatible inputs; and
• the reason a requested fusion or inference was not performed.

The user should be able to move from the scene back to the evidence, not only from the evidence into the scene.

Step 6: test the bounded behavior#

Thirty-four targeted tests passed for the ingestion, lineage, scene, and refusal paths exercised by the recorded GH1 / Bed 4 slice.

That result is deliberately narrow. It does not prove live reliability, sensor accuracy, biological prediction, safety qualification, or portability to another greenhouse. It shows that the tested reconstruction behavior performed as specified for this evidence set.

Why this sequence matters#

Agricultural systems change while data is being collected. Plants grow, light changes, equipment cycles, people move, and crop work alters the canopy. A spatial gap of minutes can be meaningful depending on the event being investigated.

The reconstruction process must therefore preserve four separations:

1. observation versus inference;
2. one capture time versus another;
3. source coordinates versus derived scene coordinates; and
4. a passed software test versus a qualified operational capability.

Those separations are what make a recorded replay useful as a foundation for a greenhouse digital twin.

Related Streamline field notes#

• What Is a Greenhouse Digital Twin?
• Why Sensor Provenance Matters in Agricultural AI
• How Streamline Detected a 43-Minute Capture Gap
• Read-Only Digital Twins Before Greenhouse Automation

Sources and scope#

This field note is grounded in Streamline’s sanitized GH1 / Bed 4 engineering case study. It omits internal paths, credentials, control-system details, and implementation-sensitive facility information. The reconstruction was recorded and read-only; it was not a live pilot.