How it works
Embertrace turns one or two camera frames into a probable origin with an uncertainty region, and uses public satellite heat to decide when and where to look. Measured inputs, modeled assumptions and limits are kept separate, because only the first kind is evidence.
Every stage evidence passes through, from a satellite pixel to a saved record.
Select a stage
NOAA publishes a fire mask, not raw radiance, so the background subtraction happens upstream. Embertrace keeps only the processed and saturated classes (10, 11, 30, 31) and reprojects each surviving pixel from the geostationary grid to a true four-corner footprint.
A frame is only half of it. The other half is where the camera was and where it was pointing at the moment of capture.
The public ALERTCalifornia feed: position, elevation, current pan, tilt, zoom and frame age. Where the feed publishes no coordinates, positions come from public CalOES and HPWREN map layers, verified against the 1,291 cameras that do publish them.
A saved or screenshotted frame has no catalog entry, so the pose comes from the footer the camera burns into the image, read by OCR in the browser.
The lowest visible point of the smoke. A vision model proposes it and a person can overrule it; a manual mark always wins. Everything downstream is that one pixel turned into a bearing.
The pixel’s horizontal offset becomes a bearing through an exact focal-length transform. What happens next depends on how much evidence there is.
90% region, along the sightlineunbounded
The field of view narrows as 2·atan(tan(FOV/2) / zoom), so at 6× a pixel carries roughly a sixth of the bearing it carries at rest. Treating zoom as a linear divisor is the error that puts an estimate a valley away.
A nine-line fan along the bearing, stepped every 100 m to 25 km and every 400 m beyond, casts the sight ray against ground height until it crosses. The ray carries a refraction term, and the crossing is refined by binary search.
Two cameras get 900 Monte Carlo samples; one camera gets 300 terrain replays. A single-camera region is long because the crossing is the uncertain part, and that shape is the honest reading of what one frame can say.
The confidence score ranks methods against each other: a two-camera fix reaches 76 at best, a single-camera bearing 30. A low number is the normal reading of one frame, not a fault, and it is never presented as a probability.
Every gate in the solve, the exact threshold it holds, and the named fire that set it — the corridor rules, the crossing vetoes, the identity ladder, and the full ledger table.
NOAA’s fire product does the per-pixel screening upstream. What remains here is history and physics: does this place’s own past explain this reading, and could the light be a reflection.
Heat seen here on other days suppresses a new observation only when it matches the pattern: same time of day, same satellites. A fire two hours outside a glint’s nightly window is not explained by it.
Power plants and landfills carry penalties by kind and distance, from live California catalogs plus a checked-in federal snapshot for the rest of the coverage area, including Nevada’s geothermal fleet.
For a single-platform daytime signal, the system computes the panel orientation that would reflect the sun into that satellite at that moment. A realistic answer lowers priority and tells the operator why. Never a veto on its own.
The Dora Fire timeline, the mirror math, the measured envelope and error rates, and what the background memory learns next.
An additive score, clamped to 0 to 100, answering how much investigation a signal is worth. Not a fire probability, and never presented as one. It carries no camera-aiming term.
| Strict thermal input | +18 | The pixel passed the strict ingestion filter. |
|---|---|---|
| Two independent providers | +24 | GOES and FIRMS both saw it. |
| Two satellite platforms | +15 | Independent support, even in the same minute. |
| Persistent heat | +12 | Three scan times spanning ten minutes or more. |
| Second scan | +8 | Same location, a distinct scan time. |
| Repeat scans | up to +14 | Three points per extra scan. |
| Sensor confidence | up to +12 | Scaled from 68% to 100%. |
| Radiative power | +1 to +10 | Logarithmic in megawatts. |
| Recency | +10 / +7 / +4 / +1 | Under 30 min, 3 h, 12 h, 24 h. |
| Single observation only | -12 | No repeat scan and no independent sensor. |
|---|---|---|
| Stale | -18 | Nothing seen for more than 24 hours. |
| Recurring background, pattern matches | -18 to -28 | Heat here on 3 or more days, and this observation fits its time-of-day and platform pattern. |
| Recurring background, pattern differs | -6 | Heat here before, but at another hour or on another satellite. History alone no longer vetoes. |
| Plausible sun glint | -6 or -12 | One platform, sun up, and a realistic panel orientation would reflect the sun into it. Stronger at glint-scale power. |
| Heat here in earlier months | -9 or -16 | The durable per-cell memory, beyond the retained fortnight. |
| Nearby heat history | -7 | Another episode close by in the retained window. |
| Known facility | down to -42 | Geothermal, biomass, landfill, fossil or solar within range. |
| Outside the United States | -12 or -40 | No camera catalog and no facility catalog cover it. Never the only thing that stops a camera check. |
Many heat pixels in one satellite image count as one scan; another distinct UTC minute counts as another. Repeat measurements keep an episode fresh, so waiting for the next five-minute GOES scan does not make it stale.
Four paths, checked most-specific first.
Each fact moves the priority score by the points beside it. Green adds. Red subtracts.
Approximate priority score28Corroborated opens at 58
One clean GOES pixel under 15 minutes old, no heat history here, no facility match, a camera in range with a usable frame.
Do not wait when a clean GOES signal is actionable. Five to ten minutes is the difference between a camera catching a fire small and catching it grown.The same fresh clean pixel carrying 25 MW or more, even where heat has recurred, as long as this observation does not fit that history's pattern.
Power is evidence about this scan, not about the neighbourhood, so it stands where the history test does not.Two or more scans, 90% confidence or better, 25 MW or more radiative power.
Bypasses the score entirely. A signal this strong is worth a third of a cent whatever the arithmetic says.Score of 58 or better, plus a repeat scan or an independent provider or platform.
The ordinary path. The threshold was lowered from 65 after a real detection landed exactly on 65 and then aged out of eligibility.A camera frame is worth buying. Ranking picks which camera, and the signal and bearing checks still decide whether anything publishes.
A newly stored thermal observation wakes the qualifier; there is no statewide camera sweep. Ineligible, stale, pattern-matched, facility-matched and duplicate episodes make no image or vision call at all, and another thermal cycle does not repeatedly score the same camera frames. The queue still lists heat no camera can reach, so an operator can know it exists.
Scored on terrain visibility along the sightline, distance from an ideal 25 km, mast elevation, frame freshness and aim. Cameras sharing a mast collapse to one; a pair is preferred when the crossing angle reaches 45 degrees.
Catalog pose is free and frames are not, so poses are watched for up to two minutes waiting for a camera to swing toward the signal before a survey frame is bought.
In daylight the model is asked for smoke. After dark a satellite-triggered check still runs, asks for glow, and must clear a stricter gate that rules out the moon and the towns by arithmetic. Assessments are rate-limited per trigger and per day.
A vision call on the retained frame, asked for the signal the sky makes possible: smoke in daylight, glow after dark, both through twilight.
The bearing implied by the model’s plume-base pixel is compared to the bearing from that camera to the thermal footprint, within the footprint’s own angular width plus the measurement uncertainties.
Satellite evidence matched within twelve hours either side of capture can carry the range while the camera carries the bearing. The fused result is labeled, and the frozen camera solve is never rewritten.
What the model is handed, the tolerance math, the night gate, and the honest limits of a single vision sample.
Embertrace is a first-detection tool. If it could read where a fire already is, it would stop measuring and start restating, and no result would be evidence of anything. None of this is available to a live solve, at any stage:
Confirmed coordinates are used in one place only: scoring accuracy after the fact, on a labeled case, with the solve already frozen. The same rule binds our own backtests, in schema and in code.
Everything Embertrace runs on, and what each piece is for. Every external service degrades rather than fails, and source freshness is shown wherever satellite evidence is.
| TypeScript on Next.js 16 | React 19, App Router, server components by default. |
|---|---|
| Vercel | Hosting, edge middleware for the access gate, and the serverless routes. |
| Postgres | Detections, revisions, retained evidence and the satellite tables. |
| Slack and email | A message when a detection publishes. Off unless configured, and the queue says which. |
| Python on AWS Lambda | Polls, reprojects and clusters. Runs on a one-minute schedule. |
|---|---|
| NOAA GOES-18 and GOES-19 | ABI-L2-FDCC fire mask from the public S3 buckets, every five minutes. |
| NASA FIRMS | VIIRS on SNPP, NOAA-20 and NOAA-21, plus MODIS near real time. |
| SQS | Hands a finished ingest to the qualifier so a camera check is never a cron guess. |
| ALERTCalifornia | Public camera catalog, current pose and latest frames. |
|---|---|
| CalOES and HPWREN map layers | Positions for the cameras the catalog publishes without coordinates. |
| Anthropic and OpenAI, via the Vercel AI Gateway | Smoke and glow detection, plume-base pixel. Optional: the app runs without it and says so. |
| Tesseract.js | Reads the burned-in footer in the browser when metadata is missing. |
| USGS 3DEP | First choice for point elevation, via the EPQS service. |
|---|---|
| Open-Meteo, Copernicus GLO-90 | Batch elevation for the terrain fan, and the first fallback. |
| Open-Elevation | Last fallback. If all three fail, nothing publishes. |
| OpenStreetMap Nominatim | Reverse geocoding for the coarse area label and the water test. |
| CEC, CalRecycle, EIA, EPA LMOP | Heat-producing facilities: California live, the rest of the coverage area from a checked-in federal snapshot. |
| MapLibre GL | Vector rendering for every map in the app. |
|---|---|
| Esri World Topo and World Imagery | Raster basemaps. Terrain is the default. |
| OpenFreeMap | Glyphs for map labels. |
Vision providers are optional and every key stays server-side. If all three elevation providers fail, the land gate cannot clear and nothing publishes.
Embertrace uses camera pixels and metadata, the public ALERTCalifornia feed, public camera-position map layers, terrain elevation, public NOAA GOES and NASA FIRMS heat detections, and federal facility datasets. It does not consult news, incident databases, dispatch, crews, media or known-fire reports.