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, including the measurement we are worst at.

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 standing on real bare earth measured by airborne lidar — USGS 3DEP QL1, 1 m posts, ~10 cm vertical, flown 2022 over a 30×30 km block of the Kern Plateau (Sierra Nevada, CA). Not a 30 m radar surface with detail painted underneath: the gullies, outcrops, road cuts and channel incision in these images are the ones that are actually there. Rothermel fire spread runs through a 3-D conifer forest; terrain and individual tree crowns cast real shadows; the whole scene is 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, sensor specs, and an in-browser instrument that runs this engine: https://ay4la.com/brasa

Validation — every stage reproduces an independent reference

Brasa's rule is that no physics stage is trusted until it reproduces something we did not make. Measured at engine a30efa8b, on the frames in this bundle:

gate	reference	this bundle	band	result
Clear-sky LWIR radiance	libRadtran 2.0.6	max |ΔTb| 0.001 K	< 0.5 K	PASS
Lidar georeference	USGS 3DEP tile corners	0.49 m	< 1 m	PASS
Fire Tb distribution	FLAME 3	within 8–11 K	—	PASS
Fire micro-texture: fill fraction	FLAME 3: 0.048	0.034	0.03–0.08	PASS
Fire micro-texture: within-fire Tb CV	FLAME 3: 0.154	0.135	0.12–0.18	PASS
Fire micro-texture: Tb p10–p90 spread	FLAME 3: 173 K	191 K	130–200 K	PASS
Fire front width (size-matched)	FLAME 3: 2.71 px	2.62 px	—	PASS
Shadow thermal lag	M4 conduction solver	reproduces	—	PASS
Rothermel spread	Rothermel 1972 / Anderson 1982	reproduces	—	PASS
Flaming residence time	Anderson 1969 (384/σ)	exact	—	PASS
Crown allometry (radius, height)	FVS Western Sierra Nevada (R5)	reproduces	—	PASS
Background texture, crown scale (4-20 m)	forest floor: 2.69 K	2.28 K	≤ 2×	PASS (0.85×)
Background texture, litter scale (< 4 m)	forest floor: 1.15 K	1.77 K	≤ 2×	PASS (1.54×)

The crowns are the stand FIA would describe — real sizes, and porous

Every conifer is DBH-sampled from the stand's diameter distribution and sized by the operational FVS Western Sierra Nevada (R5) allometry — crown radius, height and crown base one consistent tree, all published, none fitted. A DBH-60 cm overstory tree is a ~28 m conifer with a ~4 m-radius crown, not the uniform placeholder earlier versions carried. Canopy density is anchored to the reference too (Sierra mixed-conifer cover 40–70 %; ~52 % stand cover here, 91 stems/ha overstory), not to any gate.

And each crown is porous — a medium of needles, not an opaque solid — so the sun's beam is attenuated through it by Beer–Lambert, τ = exp(−G·u·L) over the path length L through foliage (G = ½, spherical leaf angle, Ross 1981; u = 1.0 m²·m⁻³, mid published conifer range). A crown is convex, so L runs from ~0 at the silhouette to the full depth on the axis: the shadow is dark in the middle and feathers over a few metres, dappling the floor rather than stencilling it. The leaf-area density lands mid-band on two stand-scale cross-checks it was never fitted against: crown LAI 4.0 (published 3–8) and sub-canopy beam transmittance 13.5 % (published closed-conifer 5–15 %).

What is known to be wrong

The litter band, at 1.54×. Below 4 m of ground scale we are half again as rough as real forest floor (1.77 K against 1.15 K). It is inside the band, and it is our weakest gate.

The cause is measured, not guessed — and the search has now cleared every easy answer. Ablation says the sub-4 m band is tree-crown shadows and nothing else (terrain shadows, painted clutter, canopy texture each move it 0.00 K). That rules out a surface-energy-balance term like latent flux (no length scale — it would scale both bands together and break the crown gate to fix the litter one). Making the crowns porous ruled out the shadow edge (1.83× → 1.57×). Smoothing the ground micro-relief ruled out the forest floor (< 0.05 K). And rebuilding the crowns to real FVS allometry — correct size, height and count — ruled out crown geometry: the litter band barely moved (1.57× → 1.54×) while the crown band improved to 0.85× of real.

That the litter band did not move under correct crowns is the finding. The leading lead — a lead, not a claim — is diffuse fill: under a real canopy, skylight and multiple scattering between crowns soften the shadow contrast, so a closed forest floor is more uniformly lit than a model casting discrete, independently-attenuated crown shadows produces. A different physics from anything tried so far, and the next thread.

How the background gate is measured. Its scale split is fixed in metres of ground and converted to pixels per frame through that frame's GSD: a filter defined in pixels measures 3.9 m of ground on FLAME 3's camera at 120 m and 6–29 m on ours at 150–900 m, so the same tree-crown shadow can fall on either side of it. And it is scored against real forested ground — FLAME 3's no-fire frames are treeless marsh, so the reference is taken from its fire frames with the fire masked out, leaving forest floor under crowns.

No detector-transfer numbers are claimed

Earlier versions of Brasa were benchmarked by training a detector on synthetic frames and scoring it on real FLAME 3 imagery (AUC 1.000 radiometric / 0.774 deployed-camera AGC, engine 0de152ab). Those numbers describe an engine that no longer exists — different ground, different sun, different embers, different labels — and they are not re-measured here. What this bundle claims is the table above.

Contents

Full bundle — brasa-wildfire-lwir-v1.0-a30efa8b-s300.zip — 99 MB (211 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)
• 228 fire (200 with visible boxes) / 72 no-fire · 115 night / 185 day · 1565 boxes
• ignition causes: campfire 58, lightning 52, powerline 59, roadside 59
• 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 reference gates above, weakest one included

Quick-look sample — brasa-wildfire-lwir-sample-v1.0-a30efa8b-s48.zip — 31 MB

48 of the bundle's 300 frames, byte-identical copies of their bundle counterparts, emitted by the same packager run and byte-compared by a live gate — so the sample you smoke-test and the bundle you train on cannot version-skew.

A note on the labels

Labels are fire regions, and the merge distance that defines a region is measured on the ground (16 m), not in pixels. This matters more than it sounds: residual embers smouldering in a burn scar sit on a ~7 m lattice, so a merge radius in pixels would group them into one scar from 600 m and shatter them into a hundred separate boxes from 200 m — an annotation describing the drone's altitude rather than the fire. Measured in metres, a burn scar labels the same way from any altitude, while a genuine spot fire, clear of the burn, stays its own object.

Reproducing

Every frame regenerates bit-identically from (engine version, seed). The bundle is stamped with the engine commit that rendered it — and the packager refuses to build from a dirty working tree, so that stamp cannot be a lie.

Reading a frame

import numpy as np, cv2
img = cv2.imread("images/train/00000.pgm", cv2.IMREAD_ANYDEPTH)  # uint16
tb_kelvin = img.astype(np.float32) / 10.0                        # deci-kelvin → K

License

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

Provided as-is, without warranty of any kind.

Superseded releases

Kept available for anyone who cited them. Prefer v1.0-a30efa8b.

• v1.0-53c1612d — the same engine with the old uniform placeholder crowns (radius 1.8–3.2 m, height 10–18 m) instead of FVS allometry. Its gate table is accurate for its own frames (litter 1.57×, crown 0.78×); a30efa8b rebuilt the canopy to real allometry, improving the crown band to 0.85× and confirming the litter residual is not crown geometry.
• v1.0-0de152ab — stands on 30 m SRTM terrain, predates cast shadows and the ground-metre label merge, and its certificate quotes the detector-transfer AUCs described above.