A national wildfire consequence model. Given an ignition, a landscape and a
day's weather, it returns a distribution of priced outcomes rather than a single
footprint.
207.0 USD M
Expected annual loss
What an average year costs, nationally. An expected value over the whole size distribution, not a bad-year scenario.
4,384,252
Hexagons, 250 m a side
The whole of continental Chile, 16.24 hectares per cell, each one carrying its comuna and región.
8,153
Recorded fires reconstructed
The mapped perimeter archive the spread model is trained and tested against, 1985 to 2018.
Seconds
Fire-level footprint, on request
An ensemble footprint priced on four terms: human safety, structures, suppression and land restoration.
The stakes
Chile's recent record
Forty seasons of CONAF's national record show a regime that changed. Not by having more
fires. By having bigger ones.
The change is in size, not count.
Between 1985–2014 and 2015–2024 the area burned per season multiplied by 2.9,
from 55,621 to 162,675 hectares a year. Over the same seasons the number of fires rose only
1.2×. The average fire went from 9.8 to 23.7 hectares, 2.4× larger. A model
that only counts ignitions cannot see this shift. You have to model size.
Two seasons carry most of that 2.9×.
2016–17 (570,197 ha) and 2022–23 (429,103 ha) hold 61% of everything burned in the
decade, and both sit far outside anything in the previous thirty years, whose worst season was
105,992 ha. Take those two out and the remaining eight average 78,432 hectares a year,
still 1.4× the 1985–2014 mean, and the decade's median is still 1.8× the earlier
median. The trend across all 40 seasons stays positive on both series: +44 fires per
year (p < 0.001) and +2,946 hectares per year (p = 0.035). The regime
did change. The 2.9× is that change plus two extraordinary years.
January 2017
~570,000 ha
The tormenta de fuego. The largest single season in the record.
February 2023
~430,000 ha
Across Maule, Ñuble, Biobío and Araucanía. 26 deaths.
February 2024
137 deaths
Valparaíso. The worst wildland-urban event in Chilean history. It anchors the fatality term in this model.
The national registry carries 109,985 georeferenced starts between
July 2002 and June 2020. Inside the modelled surface sit 10.66 million building footprints
and 16.16 million people. The ignition model's strongest predictors are temperature,
population, distance to settlement and plantation fraction.
The model
What SWIM does
01 / IGNITION
Where fire starts
A gradient-boosted classifier on hexagon-days, trained on the national registry with
deliberate burning removed. PR-AUC 0.884, ROC-AUC 0.962.
02 / SIZE
How large it grows
Final size is predicted as a continuous distribution of percentiles, not a point
estimate that the physics cannot support. Any percentile between the 50th and the 95th
can be asked for directly.
03 / SPREAD
Where it goes
A stochastic susceptible-infected-recovered automaton on the landscape raster, with a
fuel, slope and wind rate-of-spread as a susceptibility prior. Many runs give a
burn-probability field.
04 / CONSEQUENCE
What it costs
The probability field is multiplied by cost surfaces and summed over the footprint.
Four terms: human safety, structures, suppression keyed to fuel, and land
restoration.
The safety term
People are priced first, not last
Most fire models stop at property damage. SWIM carries an explicit
loss-of-life term inside the same footprint: population density under the burn-probability
field, at a fatality rate fitted from reported deaths in Chilean wildfires. It is the same
calculation that answers where it will hurt, not only what it will cost.
USD 23.9 M
of expected annual risk is loss of life
11.5%
of the national total, third of the four terms
12.6
expected deaths a year nationally, in an average year
Fire-level validation
Reconstructing recorded fires
Each recorded fire is rerun from its own ignition on its own day's conditions. The
predicted area is scored on whether the fire that actually happened falls inside it. For a
risk product the failure that matters is real fire outside the prediction, so the measure
is how much of the burn lands inside, reported next to how many times larger the predicted
area is than the burned one.
Santa Apolonia · Biobío · 26 January 2017 · 7,708 ha
It started on the worst day of the worst season in the Chilean record: the 2017
tormenta de fuego, which ended with 570,197 hectares burned nationally. Eight days of
advance on 23 km/h wind through continuous plantation. This is the case that should cost a
model the most, and it is the one it handles best: all 7,708 burned hectares fall inside
the predicted area, and it is the only fire in the set caught at the 90th percentile,
with no need to step up to the 95th. On a day of extreme wind, the size model did not have
to move up a percentile to reach reality.
Fire
Burned
Percentile
Predicted area
Predicted ÷ burned
Burn inside
Santa Apolonia
Biobío · 2017-01-26
7,708 ha
p90
24,578 ha
3.19×
100.0%
Pichaco
Biobío · 2017-01-26
5,518 ha
p95
54,497 ha
9.88×
100.0%
Alto Nilahue
O'Higgins · 2016-12-13
5,463 ha
p95
46,738 ha
8.55×
100.0%
El Carrizo
Metropolitana · 2014-01-03
14,173 ha
p95
41,062 ha
2.90×
95.7%
Tabunco / El Águila
Maule · 2017-01-26
3,215 ha
p95
6,911 ha
2.15×
92.4%
El Carrizo · Metropolitana · 3 January 2014 · 14,173 ha The largest fire on record in the Santiago region, and the biggest in this set by a factor of two. It burned for seven days on 25 km/h wind in the coastal range. This is the scale test: a 14,000 hectare fire is not the same problem as a 5,000 hectare one, because it runs long enough for the wind direction to change mid-event. The model stays with it, holding 95.7% of the burn inside the 95th percentile at 2.90× the real area.Alto Nilahue · O'Higgins · 13 December 2016 · 5,463 ha The fire that opened the 2016–17 summer, weeks before the firestorm. It is included because it shows the price of certainty: the burn falls entirely inside the prediction, but that day's 95th percentile was a wide one and the predicted area comes out 8.55× the real fire. Containing everything is easy if you draw a big enough circle, so the width of the prediction is reported alongside how much of the burn it caught.Tabunco / El Águila · Maule · 26 January 2017 · 3,215 ha The Maule coastal plantation belt, in the same firestorm week as Santa Apolonia but 150 km north and on half the wind. It is the tightest result in the set: 92.4% of the burn inside a predicted area only 2.15× the real one, a quarter of the width Alto Nilahue needed for the same containment. Where the fuel is continuous plantation and the terrain is gentle, the size distribution narrows and the answer becomes actionable.
Four fires are a demonstration rather than a statistical validation. They were
chosen because between them they cover the axes on which a spread model fails:
scale (El Carrizo at 14,173 ha), extreme wind (Santa Apolonia, on the worst day in
the record), width of the percentile distribution (Alto Nilahue, contained but loose) and
benign terrain (Tabunco, the tight case). All four were run on the same code, the same
size model and their own day's conditions, with no per-fire tuning.
The map
The national risk map
The same engine, run everywhere at once. Continental Chile is covered by 4,384,252
hexagons of 250 metres a side, 16.24 hectares each, and every hexagon carries its own answer
to one question: if a fire starts here, what should it be expected to cost in an average
year?
risk =P(ignition) × [ E(burned area) × cost density over the footprint + per-fire suppression ]
P(ignition)The chance a fire starts in this hexagon on a given day. From the ignition model, corrected so the number is a real probability rather than an artefact of how the training data was sampled.
E(burned area)How large that fire should be expected to grow. Not a single number: the average over the whole size distribution, from the roadside grass start to the megafire, each weighted by how likely it is.
Cost densityWhat the ground under that footprint is worth per hectare: people, homes, plantations and the cost of restoring the land. It varies hexagon by hexagon.
SuppressionWhat it costs to fight, keyed to fuel and to whether the fire is in the wildland-urban interface. Added once per expected fire, not per hectare.
Multiply the first three terms and add the fourth, for every hexagon on every
day of the year, and you have the map. A hexagon scores high only if fires start there,
and they grow, and there is something to lose. All three have to be true at once:
fuel with no people is cheap, people with no fuel are safe.
peak day (3 Feb 2024) against the quietest day in the daily window (14 Jan 2026). 303,346 hexes in the top loss tier against 41,834.
Checked against CONAF's own record
72,380 ha
expected burned area per year, modelled
82,384 ha
observed CONAF mean, 40 seasons · model is 88% of it
78,432 ha
typical season of the last decade, outliers removed · model is 92% of it
4,380
modelled unplanned ignitions a year, against the ~4,000 CONAF implies once intentional starts are removed
An expected value should land on the long-run average year, and this one does:
within 12% of CONAF's 40-season mean and within 8% of the typical season of the
last decade. Nothing was fitted to these numbers, the national ignition rate is pinned by the
King–Zeng prior correction and size comes from the percentile distribution. What an
expectation deliberately does not do is chase 2016–17 or 2022–23. A mean that
reproduced those seasons would be wrong about every other one; those years live in the tail
of the distribution, not at its centre.
Most starts never become fires.
The ignition model counts every registered start, median final size 0.3 hectares. The size
model is trained on fires that reached 10 hectares, mean 200. An escape probability splits
them, and the national check reproduces the observed mean size of a Chilean ignition,
14.3 hectares, rather than pricing every roadside grass start as a 200 hectare event.
