The system has three components: two indices — one daily, one seasonal — and the
satellite detection record that tests them against the ground.
Sources and standards
Index: Van Wagner, C.E. (1987), Development and Structure of the Canadian Forest Fire
Weather Index System, Canadian Forestry Service, Forestry Technical Report 35 — implementation
verified against the official Canadian system package (CFFDRS). Danger classes: the five classes of Argentina's
National Fire Management Service (SNMF), with statistical thresholds by sub-region estimated by GEOFUEGO (SNMF does
not publish its own). Live FWI meteorology: Open-Meteo (global models), each province's historical fire zones
(FIRMS 2012–2024), with 92 days of spin-up for the drought codes and the initial DC taken from ERA5; 24, 48 and
72 h outlook with the forecast meteorology of the same models. Hotspots: NASA FIRMS / VIIRS (375 m). Provincial
boundaries: IGN. Validation: 2019–2023 daily series for the Paraná Delta cross-checked against
historical detections. GeoFuego Index (GFI) 3.0: ONI (NOAA/CPC) × FIRMS detections 2001–2025
(MODIS + VIIRS, intra-sensor anomalies); drought term: DC code (CFFDRS) over the ERA5 reanalysis of
the Copernicus Climate Data Store (Hersbach et al., 2020; ERA5T in the live product), across 115 fire
zones derived from the FIRMS archive.
Antecedent: Chen, Y. et al. (2011), Forecasting Fire Season Severity in South
America Using Sea Surface Temperature Anomalies, Science 334, 787–791
(doi:10.1126/science.1209472).
The full GeoFuego Index (GFI) method is published and verifiable:
frozen calibration (3.0.0) for Argentina and Indonesia, reference rule engine, and the scripts that
reproduce every table and figure of the manuscript. The published calibration file is the one the
product loads. The manuscript — “Seasonal fire-risk skill from ENSO phase and zone-level accumulated
drought: a regional-scale categorical indicator transferred from Argentina to Indonesia” — was submitted on 12 Nov 2026 to Natural Hazards and Earth System Sciences (NHESS,
EGUSPHERE-2026-7057); it is neither accepted nor published.
3.0: ENSO signal only where the bootstrap interval of the phase ratio excludes 1; HIGH requires drought; regional extent with LOYO thresholds; the unit is the region
Open-Meteo archive (ERA5 blended with other models)
2.2.0: fixed ENSO cut-offs, Patagonia inverse, national two-thirds rule — all withdrawn in 3.0
Prospective registry: every level published since deployment (2026-09-27)
is out of sample and publicly inspectable, and is scored against the version in force at the time of its
publication. The calibration is frozen between seasons.
Another region, another country?
The method is scalable: the data sources (NASA FIRMS, ERA5, ONI) are global
and the architecture is not tied to Argentina. For each country or region, its own fire zones are derived
from the local satellite archive and the index is calibrated against its history — the same published
pipeline, with a tailored calibration. The ENSO signal on fire is documented for all of South America
(Chen et al. 2011, Science).
Already proven outside Argentina — Indonesia (Sumatra and Kalimantan):
the same method, without tuning, derived 75 fire zones and found the inverse-sign ENSO signal on its own:
with El Niño three months earlier, August–October fire reaches 7.7× its median in southern Sumatra.
The 2026 season, out of sample, came in at 7.8×. Code and data:
Zenodo v1.2.0.
Interested in the system for your country, state or region?
Write to us: hola@geofuego.com.ar
This index is a danger signal for prevention —
it is not an official alert and does not replace Argentina's National Fire Management Service
(SNMF) or civil-defense authorities. For an active fire, call 100 (Firefighters) or 911.
GEOFUEGO · satellite monitoring — a product by AgentNeo.
The daily index: Fire Weather Index (FWI)
The FWI (Fire Weather Index) is the meteorological fire-danger index of the Canadian
CFFDRS (Canadian Forest Fire Danger Rating System; Van Wagner, 1987).
It is the de facto international standard: adopted by the European EFFIS system, by the global
GWIS, and part of the same family of meteorological indices with which Argentina's National Fire
Management Service produces the national danger map. It does not detect fire: it quantifies the
moisture state of vegetation fuels and the propagation conditions — that is, the probability of
ignition and the expected fire behaviour given an ignition source.
System structure: fuel moisture codes
The system updates once a day with four meteorological variables observed at
12:00 local time — temperature, relative humidity, wind speed and 24-hour accumulated
precipitation — which drive three fuel moisture codes. Each code models a fuel stratum with a
distinct time constant:
FFMC (Fine Fuel Moisture Code) — fine surface fuels: litter, cured grass, fine
vegetal material. Time constant on the order of 16 hours: it responds to the day's weather and
governs ignition probability. DMC (Duff Moisture Code) — decomposing organic matter in the top
centimetres of the soil. Time constant on the order of 12 days. DC (Drought Code) — deep, compact organic layer. Time constant on the order of
52 days: it integrates the seasonal water deficit, so an isolated rainfall during a prolonged
drought lowers it only slightly.
This structure gives the index memory on three time scales.
The codes combine into two sub-indices — the ISI (Initial Spread Index), which integrates
the FFMC with wind and estimates the initial rate of spread, and the BUI (Buildup Index),
which integrates DMC and DC and estimates the total fuel available — and from both comes the FWI:
a dimensionless magnitude that grows non-linearly with the expected intensity of the fire front.
Computation point: the index is computed with the meteorology of
each province's historical fire zones — the points where the FIRMS 2012–2024 archive
concentrates fire activity (detection-weighted clustering; the same 115 zones of the GeoFuego Index).
The provincial light takes the worst of its zones, not an average: danger is not averaged.
The method is spatially scalable: the same engine can incorporate additional
weather stations or grid points, densifying the index down to locality or property level.
24, 48 and 72 h outlook: the same computation, fed with the forecast
meteorology of Open-Meteo's global models, for every zone and every province — the same horizon published by the
official SNMF / National Weather Service fire-danger map (the official source; GEOFUEGO complements it, it does not replace it). On the map and in the app you pick the day (Today, +24, +48, +72 h)
and everything — provinces, zones, card — is painted with that day's class. The standing value of the index is
always that of the observed day; the outlook inherits the weather model's error, which grows with lead time.
Today's data
Loading today's data…
Classification levels: the SNMF danger classes
GEOFUEGO uses the five danger classes of Argentina's National Fire Management Service (SNMF), the same
ones used by the official danger map and by the signs at the National Parks. The SNMF explains that “a given
FWI value cannot be read the same way everywhere: to compare danger across the country, thresholds are computed
statistically for each zone, and that is how the danger classes arise”. Class descriptions are the SNMF's, verbatim
(translated).
Class
What it means (SNMF)
Low
Fires that start are unlikely to stay active. However, ignition can occur near heat sources such as campfires or burns.
Moderate
Vegetation fuels are considered dry enough to sustain ignition and combustion. Although fire control is relatively easy, it can become problematic if fires are not attended to immediately.
High
Any fire that starts is a serious problem. Control becomes gradually more difficult if it is not achieved in the initial stage of fire development.
Very high
Burning conditions are critical; torching and intermittent crown fires can occur in forests.
Extreme
The situation should be considered potentially “explosive” or extremely critical.
Thresholds by sub-region (GEOFUEGO's estimate). SNMF does not publish its numbers; the
only published provincial table is Córdoba's (OHMC/IDECOR: 6 / 15 / 28 / 46). GEOFUEGO derives its own with the
occurrence threshold described in the SNMF's FWI analysis guide: each class starts at the FWI at which the
sub-region's strong fires actually occurred — Moderate, High, Very high and Extreme at the 10th, 25th, 50th and 75th
percentiles of FWI on strong-fire days (VIIRS detections above the zone's 90th percentile, static sources removed)
from 2012 to 2025, with ERA5 meteorology over the 115 zones. The 115 zones are grouped into seven sub-regions by
their FWI regime: in the Andean forest or the Yungas, strong fire occurs at values that would be routine in the
steppe or the dry Chaco.
Sub-region
Scope
Moderate
High
Very high
Extreme
NEA
Chaco, Mesopotamia and Santiago del Estero
7.0
12.5
23.7
38.4
NOA yungas
Yungas and humid valleys of the northwest
0.6
2.4
6.2
10.7
Dry NOA
Salta Chaco, dry valleys and Puna
11.5
22.3
33.2
43.5
Humid Centre
Humid Pampas and hills
5.5
11.1
19.8
31.9
Dry Centre
Monte, Cuyo and dry Pampas
10.3
19.4
31.6
46.4
Patagonia forest
Andean forest and Tierra del Fuego
0.7
4.6
16.1
23.9
Patagonia steppe
Steppe and plateau
6.8
16.8
33.3
49.5
FWI values on the ERA5 scale (start of each class). The live computation uses Open-Meteo, which
differs from ERA5 in mountain terrain: before classifying, each zone's live FWI is mapped to its ERA5 equivalent by
quantile mapping over the period in which both sources coexist (2021–2026); the card publishes the live FWI.
Out-of-sample verification (thresholds calibrated on 2012–2021, evaluated on
2022–October 2026 against strong-fire days): the “High or above” alarm captures between 78 % and 95 % of strong-fire
days depending on the sub-region, with a Peirce score (POD − POFD) of 0.25 to 0.56; the lowest is the Patagonian
steppe, where the daily FWI ranks poorly and strong fire depends more on accumulated grass. Full numbers are in the
method repository. The index remains weather-only: it does not account for fuel load, topography or the curing of
live grasslands.
Validation against fire records
Before scaling the computation nationwide, the implementation was validated in
the Paraná Delta: the daily FWI was reconstructed for 2019–2023 and cross-checked against
7,117 satellite hotspot detections from the same period.
The frequency of days with significant fire activity grows monotonically with FWI class:
3% of days in the Low class, versus 83% in the 50–70 range (Extreme). The three major
summer burning episodes of the period (December 2021, January 2022, January 2023) occurred with
FWI between 36 and 66 — all above the 95th percentile of the local historical series, whose
five-year maximum was 66.3.
The validation also revealed the index's limit. The Delta's fire regime is bimodal:
besides the meteorological summer fires, there is a late-winter mode (July–September) — burning of
frost-cured grassland — that occurs at low FWI (8–16). The days with the most detections in the whole
series (August 2021, August 2022) belong to that regime. This is a documented structural limitation of
the FWI in wetlands: its fuel model, derived from the Canadian boreal forest, does not represent cured
fine biomass that burns over wet soil. Hence the system has two axes: the meteorological index
anticipates the summer regime, and near-real-time satellite detection captures what the index does not
model.
GEOFUEGO applies that same engine — verified against the official CFFDRS
implementation — to the country's 24 jurisdictions, with the computation point described above.
The FWI expresses potential danger; hotspot detections express observed occurrence. They
are complementary magnitudes, not interchangeable ones.
Satellite hotspot detection
The points on the map are thermal anomalies detected by NASA's
VIIRS active-fire product (nominal resolution 375 m). The algorithm compares each
pixel's radiance in the mid-infrared (~4 µm) — the band where flaming combustion at
600–1000 °C emits most intensely — against its surrounding pixels; a sufficient contrast is classified
as a hotspot. Because it operates in the thermal infrared, detection also works at night and through
smoke.
A hotspot is not the same as a confirmed fire. False positives occur through
specular reflection: surfaces such as solar panels or metal roofs can reflect direct sunlight
toward the sensor, and the algorithm labels those cases as low confidence. And there are real
thermal sources that are not fires: a refinery flare, for example, produces recurrent detections at a
fixed position.
Applied treatment: low-confidence detections are discarded; detections over known static
sources (gas and oil flares, industry) are flagged and not counted: the FIRMS near-real-time feed does not
carry the archive's type classification, so GEOFUEGO reproduces it with a 910-point mask — pixels
classified by NASA as static sources in the 2012–2026 archive, VIIRS Nightfire flare sites in Vaca Muerta and cells
with recurrent fire in more than 18 distinct months — flagging every hotspot within 1 km of any of them (in Neuquén,
over 90 % of a typical day's detections). Those points remain available on the map as an optional grey layer and in
the API as fijo: 1; each hotspot is assigned to its
jurisdiction using the official IGN boundaries; and every point keeps its overpass time,
satellite, radiative power and a verification link to NASA's FIRMS viewer. Sources: NOAA-20
and NOAA-21 satellites (VIIRS sensor, 375 m), with under 3 hours of latency; S-NPP was part of the series
until 2 November 2026, when it stopped delivering data. Night-time
detections are free of solar-reflection false positives, and coincident detection by two
independent satellites constitutes cross-confirmation.
Counting criteria: the national total is the sum of the 24 jurisdictions; the satellite feed
also carries neighbouring countries and border or coastal detections with no assignable jurisdiction,
which are excluded from the counts. A declared methodological difference between the two series: the
historical archive excludes static thermal sources (industrial, volcanic and offshore); the
near-real-time series does not carry that classification — NASA adds it only when reprocessing
the data to science quality, published in the archive with a 2–3 month delay — so the recent
tail runs with a slight positive bias and is retrospectively consolidated with the archive version.
The calendar: the 2012–2025 series comes from the FIRMS archive with a single sensor
(S-NPP), which ensures year-to-year comparability. The 2026·live row integrates the satellites
in operation — three until 2 November 2026, two (NOAA-20 and NOAA-21) since then; greater detection capacity
than the archive by construction — with January–September completed from the archive.
After S-NPP, the reference of the series moves to NOAA-20, with NOAA-21 as backup: the splice is
measured on paired days of 2023–2025 (NOAA-20 / S-NPP = 0.979; NOAA-21 / S-NPP = 0.951;
harmonization method). A continuous series, with no gaps. The colour scale is not proportional to the count: it
applies a power transform (exponent 0.45) to preserve contrast in low-activity months, which a linear
scale would compress against the historical maximum.
Detection limit: a hotspot is a 375 m pixel with a thermal anomaly.
Low-intensity fires, fires under dense cloud cover or between successive overpasses may go undetected.
Absence of detection does not imply absence of fire.
The seasonal index: GeoFuego Index (GFI) · version 3.0
The GeoFuego Index (GFI) is GEOFUEGO's own seasonal index; its public reading is the
Seasonal Fire Risk scale (LOW/MEDIUM/HIGH). It complements the FWI on the time axis: the FWI
quantifies the danger of the day; this index is a categorical, regional seasonal
indicator: it says whether to expect more or less fire than normal for that month in that
climatic region. It combines two observed quantities: accumulated drought, measured in the
zones where fire historically occurs, and the phase of ENSO (El Niño–Southern Oscillation), which
enters only in the regions where its signal is measurable. The direct scientific antecedent is
Chen
et al. (Science, 2011), who proposed forecasting South American fire-season
severity from sea-surface-temperature anomalies.
Version 3.0 is the one verified in the manuscript “Seasonal fire-risk skill from ENSO phase
and zone-level accumulated drought: a regional-scale categorical indicator transferred from Argentina
to Indonesia”, submitted to Natural Hazards and Earth System Sciences
(EGUSPHERE-2026-7057; neither accepted nor published). Rules, calibration and scripts:
doi:10.5281/zenodo.24164911.
It is not an official alert nor a deterministic forecast.
The unit is the region, not the country
The GFI is regional. There is no country level and no territory level.
The unit of calibration, verification and publication is the climatic region in which the signal has a
single sign. Argentina has four: NEA (northeast, 7 provinces), NOA (northwest, 5),
Centro (centre, 7) and Patagonia (5). Indonesia has three: SUM-N (northern
Sumatra), SUM-S (southern Sumatra) and KAL (Kalimantan), with 5 provinces each.
A national aggregate of Argentina averages a winter-dry northern regime
against a summer-dry Patagonian one, with an ENSO response that changes sign between them; that is why
it is not published. The national view of the map only counts how many regions are at HIGH: it
is a count, not a level.
How the level is built
Meteorology: ERA5. The Drought Code is computed from the ERA5
reanalysis of the Copernicus Climate Data Store, at the centroid of each fire zone and in both
territories: 2 m temperature and dew point (converted to relative humidity), 10 m wind and total
precipitation; local-noon values and 24 h precipitation feed the Van Wagner (1987) equations. The
live product uses ERA5T, with about 5 days of lag, declared in every publication (date of
the latest data and days of lag). The GFI does not use Open-Meteo: a first version did, and that
precipitation series showed a break in 2017 (correlation with ERA5 of 1.00 before 2017 and 0.53–0.59
after). The daily FWI index is a separate product, with its own meteorology, described in its
section.
Fire zones. Each province is decomposed into at most five zones: the 2012–2024 VIIRS detections
are binned on a 0.25° grid and grouped with a deterministic detection-weighted k-means. This yields
115 zones in Argentina and 75 in Sumatra and Kalimantan: the places where each province
actually burns. Tierra del Fuego, Antártida e Islas del Atlántico Sur has 4 zones defined by fire
regime rather than by the automatic clustering: northern steppe and southern forest on Isla Grande,
and Isla Soledad and Gran Malvina in the Islas Malvinas (2026-11-16). The five clustered zones
counted the 2022 Tolhuin fire twice and partly depended on gas flares; and the Malvinas were missing, because
NASA's per-country archive publishes them separately. The Zenodo v1.2 package keeps the original five zones.
1 · Base. Each province's level starts at MEDIUM.
2 · Season gate. Each province's season is defined by its own archive: the rule looks at the
window of the next four months of its hotspot climatology, with weights 4-3-2-1 — the current month
weighs more than the horizon. When that window is light and the current month contributes less than
8 % of the province's annual fire, it is out of season and the level is LOW. Drought
still raises it one step: dry terrain out of season is precisely the early signal.
3 · Drought by fire zone. Each zone carries its daily Drought Code (DC, the drought code
of the Canadian system). The statistic is the 30-day trailing mean of DC on the latest day with
data, compared with that zone's threshold for that calendar month: the eleventh smallest of the 13
end-of-month values of 2012–2024. The rule therefore fires on a value that would rank among the two
highest of 13 years (top-2 of 13); its climatological rate is 2/13 = 15.4 % of
zone-months, not 10 %. If any zone of a province exceeds its threshold, the province steps up
one level and the zone that triggers it is named. At equal firing rate, this rule and a provincial
mean are equivalent in skill; it is kept because it identifies the zone responsible.
4 · ENSO, only where there is a signal. The input is the ONI (Oceanic Niño
Index, NOAA/CPC): ONI ≥ +0.5 °C defines an El Niño phase,
ONI ≤ −0.5 °C a La Niña phase. The calibration crosses the satellite detections of
2001–2025 (MODIS + VIIRS, intra-sensor anomalies) with the phase three months earlier. A region carries
an ENSO signal only if the 95 % confidence interval of the La Niña / El Niño ratio excludes
1 (year-block bootstrap). Under that criterion: in Argentina, NOA has a direct signal (more
fire with La Niña); NEA, Centro and Patagonia have none. In Indonesia, SUM-S and KAL have
an inverse signal (more fire with El Niño); SUM-N has none. In a region with a signal: a phase
opposing the signal without drought → LOW; a matching phase without drought → stays at MEDIUM
and is stated; a matching phase with drought → convergent HIGH. ENSO alone never reaches
HIGH. In regions without a signal, ENSO does not enter and the statement says so.
5 · Regional level. A region is at HIGH when the number of its provinces with drought
reaches the regional threshold, selected by leave-one-year-out (LOYO) validation to maximise the
Peirce skill score: 1 province in NEA and 2 in the other six regions. Otherwise the region stays
at MEDIUM.
Region
Provinces
ENSO signal (La Niña/El Niño ratio, 95% CI)
Regional threshold
NOA
5
direct · 1.62 (1.10–2.02)
2 provinces
NEA
7
none · 1.18 (0.98–2.02)
1 province
Centro
7
none · 1.28 (0.66–2.14)
2 provinces
Patagonia
5
none · 0.75 (0.41–1.35)
2 provinces
SUM-S
5
inverse · 0.09 (0.02–0.74)
2 provinces
KAL
5
inverse · 0.08 (0.03–0.51)
2 provinces
SUM-N
5
none · 0.56 (0.20–1.56)
2 provinces
Scope: the level is a statistical expectation, not a prediction of
individual fires. The day's danger is quantified by the FWI; operational alerting is the responsibility
of official agencies.
Presentation in the product: a declared difference from the paper
When all the provinces of a region are out of season, a regional HIGH from the
rule is shown as MEDIUM, with the label “widespread drought, out of season”. It is a
presentation decision, not a change of rule: the level given by the verified rule is still published
with every update (API field nivel_paper, alongside
tope_fuera_de_temporada), and that is the level kept in the prospective record.
Today's data
Loading the GFI state…
Classification levels
Three levels per province. Each level ships with a one-line statement declaring the driver — drought,
ENSO phase, or both — and the zone that triggers it. At the regional level the rule separates HIGH from
the rest.
Level
Criterion
Interpretation
Low
Province out of season and without drought; or, in a region with an ENSO signal, a phase opposing the signal and no drought
Less fire than normal for the month
Medium
Base level in season; or out of season with drought in any zone
Local climatology sets the expectation
High
Province in season with accumulated drought above the threshold in any zone (convergent if the ENSO phase also matches the signal). Region: provinces with drought ≥ regional threshold
More fire than normal for the month; the statement declares the driver and where
Verification: how much skill it has
The primary metric is the Peirce skill score (PSS = probability of
detection − probability of false detection): 0 is knowing no more than chance, 1 is perfect. The event
to be predicted is “detections of the month at or above the median of that calendar month”.
Full indicator, by region, out of sample. Verified at the regional level, the indicator has
positive out-of-sample skill in all seven regions. The years 2025 and 2026 entered no calibration.
Region
PSS 2012–2024 (leave-one-year-out)
PSS 2025–2026 (22 months, out of sample)
NOA
0.19
0.24
NEA
0.43
0.47
Centro
0.33
0.29
Patagonia
0.14
0.27
SUM-S
0.36
0.42
KAL
0.31
0.67
SUM-N
0.45
0.27
Drought term, by zone. In Argentina (18,096 zone-months, 2012–2024), when the rule fires,
70.1 % of cases record above-median fire in the same zone, against a base rate of 53.6 %; the
probability of detection is 20.1 % and the PSS is 0.102 (95% CI 0.062–0.137). A predictor that uses
only the fire calendar, at the same firing rate, scores 0.002. In Sumatra and Kalimantan (11,700
zone-months), the same rule gives a precision of 84.8 % against a base rate of 54.2 %, a
detection of 24.1 % and a PSS of 0.190 (0.131–0.253). Leaving one year out, the PSS is positive in
all 13 years in both territories. Out of sample: Argentina 2025, PSS 0.138; Argentina January–October
2026, 0.052; Indonesia 2025, 0.380.
Horizon. The drought term is a diagnosis of the month in progress with
a short usable tail: verified against fire one to four months later, the Argentine PSS falls from 0.102
to 0.060, 0.037, 0.025 and 0.011.
Next month: the level projected one month ahead
Besides the current month, the panel shows the GFI level projected for the following
month by province and by region. The only input that is projected is the drought term: for each fire
zone, today's Drought Code is carried to the end of next month with a regression towards the zone's
climatology (one slope per region, issue month and part of the month, fitted on 2012–2024), and the
distribution of the residuals gives the probability that the zone exceeds its drought threshold that
month. If that probability reaches 50 %, the zone enters the engine as “in drought”; the ENSO phase is
the target month's when already published, otherwise the latest published one. The combination rules, the
seasonal gate, the regional extent rule and the presentation cap are exactly those of the current month.
What is verified (2012–2026, leave-one-year-out; details in the method package): the probability
is calibrated —when the model says below 20 %, drought occurs in 4 % of cases; above 80 %, in
84 %— and the projected level matches the one the GFI ends up showing the following month in
78 % of province-months. Against fire, the projection keeps part of the current month's regional skill on
the ramp into each season (e.g. NEA April→May PSS 0.50, Central August→September 0.48, NOA
October→November 0.46, Patagonia September→October 0.32) and has no measurable skill in mid-summer or at
two or three months. That is why the panel projects a single month and presents it as what it is: the
probability that drought persists or sets in, not a fire forecast.
The panel only highlights what departs from MEDIUM: the provinces that would be
at HIGH with their drought probability, the regions at HIGH and, as a notice, likely drought in provinces
out of season. The data travels in the proximo_mes field of /api/gfi/ar.
Hindcast: the calendar × GFI
The map's calendar includes the × GFI view: for each month since 2012, the level the
system would have shown with the ONI phase current at that time and the accumulated drought of the fire
zones at that month's end, overlaid on the observed hotspot count. Over 2012–2024 it is an exercise in
coherence over the calibration period, not independent verification; the out-of-sample
verification is the one in the previous section, and the record since deployment tests the index
prospectively, season by season.
Scalable: the same method, tested in Indonesia
The GFI is not tied to Argentina: fire zones, seasons, drought thresholds and ENSO
signals are derived from each territory's own data. The test was to run it without tuning on
Sumatra and Kalimantan (Indonesia): 75 fire zones, and the method found on its own an ENSO signal of
the opposite sign to northwestern Argentina's — with El Niño three months earlier, August–October fire reaches
7.7× its median in southern Sumatra and 4.9× in Kalimantan. The 2026 season, which entered
no calibration, came in at 7.8× and 5.7×.
Manuscript submitted, not peer-reviewed. The manuscript that verifies version 3.0 has been submitted to
the journal; its results have not yet passed peer review.
Argentina 2026: almost no fire. The 2026 season was exceptionally quiet: under a strong El Niño
(ONI 2.46 for August–October), July–October detections were 0.41 of the median nationally and 0.22 in
NOA. With so little fire, the skill of any indicator is bounded by the absence of events: it is the
limiting case, not a failure or a success. At the finer scale (province-month), out-of-sample
verification is positive in NEA (0.11), Patagonia (0.08) and the three Indonesian regions (0.34 to
0.51), and negative in NOA and Centro.
Patagonia. An earlier version declared an inverse ENSO signal in Patagonia. It does not meet
the current criterion (ratio 0.75, interval 0.41–1.35, which includes 1) and it failed out of sample:
January–March 2026, under La Niña, recorded 1.54 times the median. In 3.0 Patagonia carries no ENSO
signal and its level depends on drought alone; it is the region with the lowest skill over 2012–2024
(0.14).
The country is not a valid unit. The national aggregate of Argentina has no out-of-sample skill
at the province level, and the ENSO term at country scale has a seasonal PSS of −0.01. That is why the
GFI publishes no country level.
The ENSO term is small and short-sampled. Its seasonal skill is low everywhere (PSS 0.12 in
NOA, 0.19 in SUM-S, 0.21 in KAL, over 25 seasons): it identifies a few seasons with high confidence and
is silent in the rest. In NOA it rests on three La Niña seasons. None of the 50 tests of the lag sweep
remains significant after Bonferroni correction; the declared signals rest on the bootstrap
intervals.
Detection and cloud cover. The verified event is a detection count, and drought reduces cloud
cover, which improves detection independently of how much burns. Part of the drought term's skill may
reflect better observability; separating the two is declared work.
Short climatology. With 13 years, the drought threshold can only be resolved by ranks; it
improves with every added year.
Meteorology lag. ERA5T arrives with about 5 days of lag: the level reflects the drought
accumulated up to that date, which is declared in every publication.
Hydrological amplifiers. Episodes like the Paraná low-water stand — which leaves the
wetland's fuel exposed and continuous — belong to the index's next layer: integrating them with the
corridor's hydrometry is the declared line of work.
Resolution in wetlands. The DC models the forest soil of the
Canadian system, and the meteorology comes from reanalysis cells of about 25 km. In wetlands and
peatlands, the value works as a regional drought tendency — a useful approximation, read alongside the
satellite detection that captures what the model does not represent.