Calibration strategy - 18 May 2026
Make every thermal alert auditable
This page specifies a field-reference workflow for turning raw thermal frames into calibration-flagged evidence rather than untraceable operator imagery. It is a written procedure awaiting hardware; no bench or field calibration has been performed yet.
Field kit
The payload needs reference targets, not just a factory-calibrated camera.
What changed
Calibration is part of the field workflow and validation boundary.
The strategy uses two reference anchors, detector warm-up, shutter/NUC logging, window-state checks, and repeated reference passes when the route allows them.
Frames can be benchmark-grade, advisory-grade, or uncalibrated. That status follows every alert and dataset export so later validation cannot hide stale calibration.
Flight cadence
Reference checks bracket the survey and repeat when drift risk increases.
| Stage | Minimum action | Why it matters |
|---|---|---|
| Bench characterization | Record warm-up, NUC/shutter behavior, two-reference response, and window-in/window-out offset. | Prevents the first field flight from discovering detector behavior after the data is needed. |
| Pre-flight stabilization | Use a 20 minute detector warm-up as the conservative starting rule unless bench data supports less. | Microbolometer drift is an operating condition, not a bookkeeping detail. |
| Launch reference pass | Capture raw or radiometric frames of warm and ambient targets before survey frames are scored. | Creates the first calibration segment for the flight manifest. |
| In-flight repeat | Add short reference waypoint passes when route geometry, elapsed time, or payload temperature change justifies them. | Repeated references make drift correction defensible instead of guessed. |
| Survey frames | Attach calibration version, reference age, detector settings, NUC state, window state, and payload temperature. | Every alert remains replayable against the thermal source data. |
| Post-flight closeout | Repeat reference frames, inspect the window, and mark the flight benchmark, advisory, or uncalibrated. | Validation metrics cannot be reported before the calibration state is known. |
Quality gates
The first calibration gates are provisional but enforceable.
Benchmark-grade
Reference residual within 2 deg C after correction, with stable NUC state and known window condition.
Advisory-grade
Reference residual within 5 deg C, or only pre/post-flight references available. Useful for review, not headline metrics.
Uncalibrated
Missing reference frames, stale correction, unknown shutter/NUC state, or contaminated window state.
Still open
The final absolute radiometry budget remains part of detector, target-signal, and atmosphere closure.
Metadata contract
Calibration state has to travel with raw frames, alerts, and dataset exports.
Each flight manifest should record detector branch and serial, optics profile, window/filter state, firmware and software versions, weather source, route, operator, and storage root.
Each reference record should preserve target ID, target type, emissivity assumption, measured temperature, frame IDs, GPS position, capture time, and validity flag.
Each alert packet should include calibrated and raw values, threshold provenance, calibration version, reference age, detector settings, and a validation-exclusion flag when the frame is only advisory-grade or uncalibrated.