The wildfire
intelligence layer.
Turn any landscape, fire model and observation stream into a living, auditable wildfire twin.
Live on screen: CASE 001 — La Granja de San Ildefonso, Segovia, 4 August 2019. Real terrain, real fuel, real wind field, real detections.
Not another fire map
Traditional systems show where the fire is.
FireTwin models what the fire is becoming.
A map is a photograph of the present. A twin is a state with a past, a physics and a distribution over what comes next — and it can be asked why it believes what it believes.
FireState — the object underneath
Arrival time T(x,y)
Minutes from ignition to every cell, on a fixed 20 m grid. Not a polygon per frame: one surface the whole timeline is read from.
Active front
The band where T(x,y) equals the clock. The front is derived from the state, never drawn by hand.
Uncertainty
An ensemble of members over six axes, not one deterministic line. What survives the observations is a distribution.
Provenance
Every object carries its class — official, observed, derived, reconstructed, simulated — plus source, timestamp and confidence.
These are not decorative colours. They are the classes, and they are the same ones the viewer reads from the case manifest — so a reconstruction can never be mistaken for a record.
Evidence Graph
Classes say what an object is. The graph says what it rests on.
Every object is a node, and the edges are the relations that matter when a figure is questioned: what it derives from, what supports it, and what contradicts it — because in this case five burned-area figures contradict each other and the record keeps all five. Ask the twin why it believes the fire reached a ridge at a given hour and the answer is a path through that graph, ending in a satellite pass with a timestamp, not in a confidence score.
The FireTwin stack
Four things, in this order
-
01
RECONSTRUCT
Rebuild a past fire from the official record, the terrain, the fuel and the physics — with every layer labelled by where it came from.
-
02
OBSERVE
Fold in satellite detections, official perimeters and station data as evidence with timestamps, not as decoration.
-
03
BENCHMARK
Score the model against those observations with a metric that knows suppression happened — and publish the score whether it flatters the model or not.
-
04
PREDICT
Only then run forward, as an ensemble, and report bands with the hypotheses attached.
The order is the product. A prediction engine that has never been benchmarked against a reconstructed fire is a plausible-looking picture, and there are plenty of those already.
Model Bus
One contract. Any engine.
The temptation with a single engine is to write a wrapper. It works once, and you throw it away when the second engine arrives — because by then half the pipeline speaks the first engine's dialect. So the contract came first. This is the real one, the one in the repository.
Diagram: normalised inputs enter a RunSpec, any adapter runs its engine, and a RunResult comes back with arrival time, perimeters and provenance. Observations enter separately, at the scoring step.
Normalised inputs
- terrainelevation model, working grid
- fuelScott & Burgan classes, dated
- weatherdownscaled wind fields + fuel moisture series
- ignitiontime, position, initial radius
- paramsthe six ensemble axes + seed
The contract
RunSpec → adapter.run() → RunResult
- the version comes from the binary, not from a constant
- validate() fails before burning CPU
- same spec → same raster, or the seed is declared
- a failure returns an error, never half a raster
- no engine dialect escapes upwards
Normalised outputs
- arrival_timeminutes since t0, float32, NaN where the fire never came, on the exact input grid
- perimetersone feature per dump step, with its minute and its area
- intensitywhere the engine reports it; declared as absent where it does not
- metadataclass, source with engine version, timestamp, effective parameters, diagnostics
Where uncertainty attaches
A single run comes back with confidence: null, on purpose, and with the reason written into the file: one run has no confidence. The ensemble has it, measured against observations. Putting a number there would have been inventing it.
Where observations enter
Not into the run — into the score. The engine propagates without knowing that aircraft existed; the scorer is the one that knows. Keeping those apart is what makes the benchmark below mean anything.
Engines on the bus
Loading the engine roster…
Benchmarking
Scored against what actually happened
Every figure in this section is recomputed by a script that re-runs the calibrated member and re-scores it against the observations. None of it is typed into this page. If the script has not run, the section stays empty rather than showing a number nobody can reproduce.
Symmetric scoring would be a lie
The observed fire is the trace of a fire that had roughly 1,395 drops thrown at it. Scored on raw overlap, the winning member is the one that spreads worst — the one that puts itself out. So overshoot is weighted by the response that was overhead at that hour, and the weights are declared, not tuned.
A detection proves arrival. Its absence proves nothing
If a satellite sees no fire at 03:27 it may mean there was none, or that the pass did not cover it, or that cloud hid it. So arriving late at a detected hotspot is penalised hard, and arriving early is penalised gently.
Loading the measurements…
Cases
CASE 001 · La Granja de San Ildefonso
Segovia, Spain · 4–21 August 2019 · file FIRE-2019-0400089. Reconstructed in the style of an accident investigation: canonical figures only, each with its source and its class, and where the record disagrees with itself, every version is kept.
| Item | Value | Class |
|---|---|---|
| Burned area (official register) | 385.77 ha | OFFICIAL |
| Copernicus EMSR375, 6 Aug | 370.2 ha | OFFICIAL |
| Own dNBR perimeter | 352.0 ha | DERIVED |
| Detection | 4 Aug, 14:46 | OFFICIAL |
| Ignition time (median, robust 95% CI) | 14:36 · 13:47–14:45 | RECONSTRUCTED |
| Satellite detections in the window | 80 · 6 passes | OBSERVED |
| Point of origin | as declared in the report | OFFICIAL |
Five different burned-area figures coexist in the record and none of them is “the right one”; all are kept and labelled. The inferred ignition interval covers 58 of the 61 minutes it was searched over — it confirms the window and tilts it, it does not narrow it. FireTwin reconstructs when, where and how. Never who.
CASE 002 IN RECONSTRUCTION
Selection criterion: observational richness — a Copernicus EMS activation with more than one delineation, so the perimeter has a time series and not just an ending.
CASE 003 IN RECONSTRUCTION
Statistical learning stays gated until three fully observed fires exist, split by whole fire — never with one fire on both sides of the split.
Why FireTwin
Five commitments, each with a cost we accepted
Model-agnostic
The engine sits behind a contract, so it can be replaced. The cost: writing the contract before we needed it.
Evidence-based
Objects link to what they derive from and what contradicts them. The cost: declared gaps where a number would have looked better.
Suppression-aware
The scoring knows an operation was overhead. The cost: our headline overlap number looks worse than a naive one would.
Uncertainty-first
Intervals ship next to every figure. The cost: publishing an interval that stayed wide because the evidence was thin.
Digital-twin native
One canonical state and one clock, shared by the viewer, the metrics and the narration. The cost: no shortcut renders that the data cannot back.
Who it is for
Fire agencies
Post-incident reconstruction, training scenarios, and a state that can still be interrogated years later.
Land & forest managers
Fuel state over time, treatment counterfactuals, exposure of stands and infrastructure.
Utilities & infrastructure
Arrival-time estimates at lines, substations and access roads, each with its interval.
Insurers & climate risk
Physically grounded event reconstruction for loss attribution and scenario work.
Research
An open contract, a published scoring method and a case whose numbers can be recomputed from the repository.
Decision support for qualified wildfire professionals. FireTwin does not replace judgement, and every figure it shows carries its interval.