AY4LA/brasa-wildfire-lwir

Brasa — synthetic wildfire LWIR thermal imagery (v1.0)

Labeled synthetic thermal imagery for training and testing wildfire-detection models — shipped with the evidence that its physics is real.

Real labeled wildfire thermal data is scarce: fires are dangerous to instrument, aerial campaigns are expensive, and "ground truth" is usually a threshold drawn on the very pixels a model trains on. Brasa is the synthetic alternative. Every frame is a fully simulated wildfire on a real mountainside — SRTM terrain, Rothermel fire spread, a 3-D conifer forest — imaged through a physically modeled thermal camera (optical PSF, NETD, fixed-pattern noise, ADC). Labels are projections of the simulated ground truth: pixel-exact, never thresholded from the image.

Full release dossier, validation figures, and sensor specs: https://ay4la.com/brasa

Does it transfer? (measured, with the caveat inline)

A YOLO detector trained on nothing but Brasa frames — it never saw a real thermal image — evaluated on 738 real FLAME 3 wildfire frames (Sycan Marsh prescribed burns, aerial radiometric thermal):

domain	AUC	note
Radiometric (calibrated Tb → 8-bit)	1.000	every real fire frame outscored every no-fire frame at zero false positives. Saturated: a plain temperature threshold also scores 1.000 here — this proves transfer, not ML advantage.
Deployed-camera AGC 8-bit video	0.774	the honest frontier: 0.384 TPR @ ≤5% FPR, from 0.079 one engine milestone earlier.

Full caveats (including why FLAME 3 cannot score open grass/brush transfer) ship inside the bundle's validation certificate and on the release page.

Contents

Full bundle — brasa-wildfire-lwir-v1.0-0de152ab-s300.zip — 90.3 MB (212 MB unpacked)

• 300 frames (270 train / 30 val): 16-bit radiometric PGM (deci-kelvin — pixel/10 = Tb_K, so fire cores are represented, not clipped)
• Labels: YOLO and COCO (category fire), with physical ground truth riding along — FRP, fire area, plume height, peak Tb, contrast. No-fire frames ship as labeled negatives.
• Provenance manifest — every knob that generated each frame
• 24 false-colour previews
• Validation certificate — live-run integrity gates (determinism, coverage, label sanity, sample-sibling) + the engine's reference gates (radiometry vs libRadtran 2.0.6, fire micro-texture vs real FLAME 3 bands, Rothermel spread reproduction)

sha256: CA60E40C8AF1C685C31E2822ADA8C00526ABD5C0F0E68231C886F22115066CEB

Quick-look sample — brasa-wildfire-lwir-sample-v1.0-0de152ab-s48.zip — 27.7 MB (75 MB unpacked). 48 of the bundle's 300 frames (41 fire / 7 no-fire, 20 night / 28 day, 826 boxes), byte-identical copies of the bundle's own files, emitted by the same packager run: same schema, same labels (COCO and manifest keep the parent's ids), same license, a preview for every frame. Smoke-test on the sample, then train on the bundle — a live certificate gate byte-compares every sample file against its bundle counterpart, so the two cannot version-skew.

sha256: 2B6E6DA5928431979756BABFD8A23D01BB16B55E7DF05895CE22B0543E0041ED

Provenance & determinism

Engine commit 0de152ab. Every frame regenerates bit-identically from (engine version, seed). No physics stage is trusted until it reproduces an independent reference — the certificate in the bundle records the gates this build passed.

License

CC BY-NC 4.0 — free to use, share, and adapt for research, education, and other noncommercial purposes, with attribution ("Brasa — ay4la.com"). Commercial use — including commercial deployment of models trained on this data — requires a separate license: elroy@ay4la.com.

Cite as

Brasa — synthetic wildfire LWIR thermal imagery, v1.0 (AY4LA, 2026).
https://ay4la.com/brasa