Fire radiometry synthesis - 2 June 2026
The fire literature makes mixed-pixel radiometry the first detection gate.
EmberScope can keep the first payload on the uncooled LWIR path, but detection evidence has to preserve raw thermal frames, fill fraction, calibration state, and false-positive context.
Design decision
LWIR remains the first build path, but display imagery is not enough evidence.
Evidence read-through
The papers agree that small fires are diluted radiance signals before they are detection labels.
| Evidence | What it shows | What EmberScope should do next |
|---|---|---|
| Dozier subpixel retrieval | Thermal pixels mix target and background radiance; target area fraction is part of the measurement. | Keep fill fraction, apparent contrast, and two-pixel energy in every performance gate. |
| Burke drone peat-fire trials | Drone LWIR can see hot buried or open fires through smoke, but the demonstrated targets were much larger than centimetre-scale hot spots. | Use the result as feasibility support, then stress-test weaker and smaller targets. |
| Allison airborne review | TIR helps through smoke; hotter fires shift toward MWIR, and reflected or sun-heated backgrounds create ambiguity. | Keep LWIR for the first payload while preserving MWIR or dual-band retrieval hooks. |
| Sousa thermal-camera pipeline | AGC, color palettes, NUC events, and saturation can change display contrast without a matching scene change. | Store raw/pre-AGC frames, detector state, calibration state, and display transform metadata. |
| Wang small/cool fire algorithm | Global thresholds miss regional small fires; smoke and contextual tests recovered omitted cases. | Make thresholds altitude-aware, background-relative, multi-frame, and reviewable. |
Implied requirements
The detection stack needs measurement provenance, not just a hot-pixel trigger.
Raw thermal source
Capture raw or radiometric frames before display compression, AGC, and color mapping change the measurement surface.
Detector state
Log detector temperature, NUC events, saturation range, lens/window/filter state, and calibration references.
Mixed-pixel metric
Report target fill fraction and optical ensquared energy beside apparent temperature contrast.
Contextual threshold
Adapt thresholds to background distribution, altitude, GSD, dwell, smoke, and local clutter instead of one global cutoff.
Bandpass growth path
Hold MWIR or dual-band sensing as a retrieval and classification upgrade, not a blocker for first LWIR detection evidence.
Stakeholder values
Close acceptable miss rate, false-alarm rate, minimum target size, target class, survey altitude, and smoke/weather envelope.
Validation route
The first useful field evidence should target the weak cases.
The next validation packets should start with bench raw-frame checks at known fill fractions, then move to 1 cm, 3 cm, and 10 cm hot-spot surrogates in low-altitude flight geometry.
No-fire rural clutter, smoke or haze, sun-warmed objects, burned ground, and hot vehicles or metal should be treated as required false-positive classes before detection claims are made.