Brasa — synthetic wildfire
thermal imagery.

No black box.

The “measurement” is not a filter over a picture — it's a chain of named, individually-validated effects. Watch one fire frame lose exactly one thing at a time: optical blur (Airy PSF), detector sampling, 40 mK temporal noise, fixed-pattern noise, then the 14-bit ADC. The final panel is bit-identical to the delivered frame.

Validated against independent references.

No physics stage is trusted until it reproduces an independent reference. Each figure carries its own method box — reference, metric, threshold — printed by the same code the validation gates run on. In the table below, each gate is tagged by what it is checked against: real data, a reference code, or an analytic law.

Does it transfer to real fire?

The test that matters: train a detector on Brasa frames only, then score it on 622 fire + 116 no-fire real FLAME 3 frames the detector never saw. It is run as a matrix — 2 grounds × 3 scene conditions, 6 arms in all — because a single number would hide what the answer depends on. Two sensor domains separate, and they answer differently.

how to use it — the split above is the guidance. If your camera hands you calibrated radiance, the radiometric column is saturated and Brasa is training data. If it hands you 8-bit auto-gained frames — which is most deployed cameras — the AGC column is where you actually are, and Brasa is pretraining and augmentation data: pretrain on Brasa, then fine-tune on a small real set. Nothing measured here supports replacing real data.

The two grounds are not a controlled comparison: the lidar block yields 33 scenario windows against 244 for SRTM, so terrain fidelity and scene diversity move together here and cannot be separated from these runs. Read the rows as two worlds measured, not as a ranking of grounds. frame-level ROC against FLAME 3's own Fire/No-Fire labels; frame score = max detection confidence. No box-derivation convention anywhere near the headline number (the Track C confound lesson).

provenance — engine 73349c60, one commit past the a30efa8b the gate table above was measured at; that commit adds the ablation switches this matrix needs. Some arms carry a `-dirty` stamp in their raw arm.json: the working tree changed during the run (documentation and analysis tooling), while engine/ did not. Frames are a pure function of (engine code, seed), so those arms are reproducible from this commit; the stamp was over-conservative and its scope has since been corrected. Checked rather than asserted: seeds regenerated from a dirty-stamped arm and byte-diffed against the originals — bit-identical.

the open question — a physically-modeled AGC stage inside the synthesis loop, so a detector trained purely on synthetic frames transfers to real deployed cameras

Three reference sensors.

Every number below is computed from the sensor model — none typed by hand. These three are reference profiles that bracket the model's range — GSD 0.82→13.6 m/px, pixel-limited→diffraction-limited optics: one resolves the fire, one sees it sub-pixel. The v1.0 bundle's own frames are rendered across 150–900 m altitude, not off these exact three. Your sensor spec is a profile away.

The honest part.

A number without its caveats is marketing. These ship with every bundle, alongside the validation certificate:

• ›The litter-scale background band runs 1.54x rougher than real forest floor. It passes, and it is where we are weakest. The cause is measured -- crown shadows -- and the next lead is named above.
• ›Two engine stamps on this page, kept apart rather than merged. The gate table was measured at a30efa8b; the sim-to-real transfer matrix and the radiometric distribution gap at 73349c60, the next commit, which adds the ablation switches that matrix needs. The distribution figures changed materially when they were re-measured -- background gap 7.8 -> 6.6 K, fire-to-background contrast 3.5 -> 4.7 K -- and the cause is staleness, not regression: they had not been re-emitted since before M24 replaced a fixed downwelling-longwave constant with a temperature-dependent sky, so the old values described an engine five milestones behind the one they were published beside. The background cooled toward the real reference, which is an improvement. Contrast is fire minus background and the fire did not move at all, so contrast widened by exactly the same 1.2 K -- the metric got worse because the physics got more right.
• ›The background and contrast figures pool day and night frames, and those two populations sit about 31 K apart, so the pooled median is sensitive to how many day pixels survive the 330 K cut that defines background -- a cut that selects on the very quantity it then reports. Measured separately, our night background sits within about 2 K of the real reference and our day background runs about 34 K warmer. Whether that is a daytime defect or a comparison with no counterpart in the reference depends on FLAME 3's acquisition conditions, which we have not confirmed. Reporting this gate split by day and night, and defining background by a non-fire mask instead of a temperature threshold, is open work and is not done.
• ›Labels are fire REGIONS. Residual smouldering hotspots inside a burn scar merge into their scar at a 16 m ground distance rather than each becoming an object; a genuine spot fire, clear of the burn, stays its own box. A positive box marks an active-combustion region -- flaming or still-smouldering -- not a new ignition, so a warm burn scar with live embers is a true positive, not a false one.
• ›FLAME 3 measures one world (forested prescribed understory burns), so open grass and brush fire behaviour cannot be scored against it -- only asserted from Rothermel.
• ›NETD is quoted at 300 K; the 14-bit ADC is linear in radiance over a fixed 260-900 K span (headroom so fire cores are not clipped), so near ambient the quantization step is coarser than the NETD -- the NETD is the detector spec, not the delivered background floor.
• ›The sensor chain is static-platform -- optical PSF, detector sampling, NETD, fixed-pattern noise (PRNU/DSNU/column), then the 14-bit ADC -- with no motion or scan smear modelled, so along-track resolution on the moving airborne and long-range profiles is optics- and detector-limited only.
• ›Background warm clutter is modelled -- solar-loaded roads and waterways, terrain insolation, and material-dependent surface temperature -- but no non-fire source is driven to fire-adjacent apparent temperature, so frame-level fire/no-fire is near-separable by absolute temperature (this is why the radiometric AUC saturates at 1.000, a baseline rather than a result). A hard negative at ember-or-fire brightness -- a sun-baked rock or hot artefact pushed into the fire Tb range -- is a next scene addition, not a solved case.

deterministic provenance: every frame regenerates bit-identically from (engine version, seed). the numbers on this page are imported from the engine's machine-derived results files — never re-typed.