Wildfire Exposure Situation Report , Iberian Peninsula: Posterior-weighted expected exposure, triangulated against authority and press reporting

Capybara Center for Risk Analysis · situation report · under JANUS

Wildfire Exposure Situation Report , Iberian Peninsula

Posterior-weighted expected exposure, triangulated against authority and press reporting

HazardWildfire, structural & short-horizon exposure

AreaPortugal (Run A); Iberian Peninsula (Run B)

Run windowEarly July 2026 (execution-time run)

HorizonGFS short-term, ≤504 h (~21 days)

Provenance. Legacy Capybara Bayesian wildfire model (Google Earth Engine). Current fields key off execution time: GFS accumulation, CHIRPS trailing window, and the FIRMS active-fire overlay (last 7 days). ENSO table (from code): el_nino = 2019, 2023, 2024; all_years = 2019–2025. Config: GFS ≤504h · fineScale=500 · exposureScale=1000 · profile=standard. Code as FDO: the legacy script will be archived as a citable release; every claim below carries derivedFrom: capybara-legacy [PID pending].

Legend modeled risk: low → high (grey → amber → orange → dark red) FIRMS active fire, last 7 days (observed, separate layer)

Reading correction. The bright yellow blobs on these maps are not the top risk class. They are FIRMS satellite detections of fire actually burning in the seven days before the run. Dark red is the highest modeled-risk class. This report treats red as modeled risk and yellow as observed fire, and triangulates both against independent reporting.

The model concentrates wildfire exposure in northern and north-central Portugal and across northern and north-eastern Spain. Independent authority and press reporting from the same window corroborates the picture: Portugal declared a nationwide state of alert (3–6 July), and active fires are burning in Leciñena (Aragón) and at the Gavarres massif (Catalonia), exactly where the model's FIRMS detections and high-risk cells fall. The figures are scenario estimates of exposure, and the corroboration below establishes consistency, not predictive skill.

§0Model and configuration

Both runs are posterior-weighted expected-exposure surfaces from a Bayesian model that fuses a VIIRS fire-density and satellite-embedding prior with a GFS-driven forecast hazard, crossed against ten asset layers. This is the standard forecast profile. A prior-dominant structural run, which down-weights short-horizon weather in favour of persistent terrain and fuel susceptibility, is the appropriate companion for siting durable physical intervention and is recommended as a follow-up.

§ARun A · Portugal, El Niño analog years (2019, 2023, 2024)

Where it concentrates. A broad band across the northern and north-central interior, from the Marco de Canaveses–Amarante corridor through Vila Real into the Beira Alta highlands around Castro Daire, São Pedro do Sul and Sernancelhe, with a prominent cluster near Águeda and Albergaria-a-Velha. Lisbon, Setúbal and the southern Alentejo and Algarve are largely outside the flagged surface.

Asset	Share	Expected	Total
Population	6.8%	~568,000 ppl	8,321,886
Protected areas	9.2%	~185,400 ha	2,007,611
Built-up footprint	9.1%	~805,600 ha	8,874,044
Electric grid	11.0%	~2,460 km	22,473
Roads	10.1%	~3,293 km	32,622
Wind farms	23.9%	49 sites	203
Cropland	3.0%	~42,100 ha	1,404,392
Urban area	1.7%	~2,283 ha	132,382
Solar farms	2.2%	1 site	61

Triangulation · corroboration

Portugal's Autoridade Nacional de Emergência e Proteção Civil (ANEPC) warned of extreme rural-fire danger, worst in the Norte and Centro regions, from 2 July, and the government declared a state of alert for all mainland Portugal from 3 to 6 July (portugal.gov.pt; Observador). The IPMA placed almost the entire country at very-high-to-maximum fire danger, with only six coastal districts at moderate, and over 1,000 firefighters were deployed (Euronews). The north-and-interior gradient in the official danger picture matches the model's northern concentration, and ANEPC itself noted the most-exposed areas coincide with territories of historically high fire incidence, which is the same signal the model's fire-density prior encodes.

{ "@id": "nanopub:tri_PT_norte_2026-07",
  "assertion": { "subject": "model high-risk surface, N & Central Portugal",
    "predicate": "corroboratedBy",
    "object": "ANEPC state of alert, Norte/Centro extreme danger, 3–6 Jul 2026" },
  "provenance": { framing: "consistency, not predictive skill",
    source: "portugal.gov.pt, 2026-07-02/03", ensoPhase: "el_nino",
    derivedFrom: "capybara-legacy [PID pending]" },
  "pubInfo": { "identifier": PID pending, "license": "CC-BY-4.0" } }

Notable finding. Wind farms are the most disproportionately exposed asset in this run, nearly one in four flagged, consistent with siting on the northern interior ridgelines that coincide with the flagged highlands. Linear infrastructure, the grid and roads, is flagged at roughly one tenth. Worth routing to grid and renewables operators, not only to civil protection.

§BRun B · Iberian Peninsula, all-years baseline (2019–2025)

Where it concentrates. Along the northern rim, Galicia and the Cantabrian fringe, the Basque Country and Navarra, and the Pyrenean foothills, with detail clusters in Aragón (Leciñena, in the Monegros) and Catalonia (the Guilleries–Montseny massif inland, and the coastal Baix Empordà near Palamós). Interior and southern Spain are largely outside the flagged surface.

Asset	Share	Expected	Total
Population	1.6%	~552,000 ppl	33,597,289
Protected areas	3.2%	~454,200 ha	14,243,109
Built-up footprint	2.5%	~1,241,300 ha	50,564,785
Electric grid	2.5%	~2,163 km	87,388
Roads	2.6%	~3,640 km	143,022
Wind farms	4.7%	16 sites	331
Cropland	0.9%	~122,800 ha	13,001,583
Urban area	0.6%	~3,226 ha	558,677
Solar farms	1.1%	0 sites	28

Triangulation · Aragón (direct hit)

Two wildfires active this week in Leciñena (Zaragoza) and La Fueva (Huesca) had burned close to 3,000 hectares of high-value forest, cutting the A-129 and forcing an evacuation at Morillo de Monclús, with AEMET warning of a fresh heatwave above 40°C (Moncloa). Leciñena is precisely where the model places a flagged cluster and a FIRMS detection, so the observed fire coincides with both the modeled risk and the active-fire overlay.

{ "@id": "nanopub:tri_ES_lecinena_2026-07",
  "assertion": { "subject": "model-flagged cell + FIRMS detection, Leciñena (Zaragoza)",
    "predicate": "coincidesWith", "object": "active wildfire ~3,000 ha (with La Fueva)" },
  "provenance": { framing: "concurrent observation, consistency",
    source: "moncloa.com, 2026-07-03", derivedFrom: "capybara-legacy [PID pending]" },
  "pubInfo": { "identifier": PID pending, "license": "CC-BY-4.0" } }

Triangulation · Catalonia

A serious forest fire at La Bisbal d'Empordà reached the Gavarres massif and had scorched more than 1,300 hectares, driven by the tramuntana wind and a prolonged heatwave, with fire danger high or very high across many Catalan counties (Ara). The Gavarres sit in the coastal Baix Empordà beside Palamós, where the model's FIRMS detection falls, so the observed fire corroborates the coastal Girona flag.

Divergence · logged

The active Catalan fire is coastal (Gavarres, near Palamós, the yellow FIRMS cell). The model's separate inland red cluster in the Guilleries–Montseny massif has no confirmed active fire in this window: it is modeled structural risk that has not been realized. This divergence is recorded rather than smoothed, and it is exactly the kind of case that later outcome tracking must resolve.

{ "@id": "nanopub:tri_ES_gavarres_2026-07",
  "assertion": { "subject": "FIRMS detection, coastal Baix Empordà (near Palamós)",
    "predicate": "coincidesWith", "object": "La Bisbal d'Empordà / Gavarres fire, >1,300 ha" },
  "provenance": { framing: "consistency; inland Montseny flag unrealized",
    source: "ara.cat, 2026-07", derivedFrom: "capybara-legacy [PID pending]" },
  "pubInfo": { "identifier": PID pending, "license": "CC-BY-4.0" } }

§CReading the two runs together

The higher exposure fractions in Run A look like an El Niño effect, but the two runs cannot establish it: Run A covers Portugal alone while Run B covers the whole peninsula, so the denominators and geography differ, and part of the gap is simply that Portugal is smaller and more fire-exposed than the Iberian average. A clean comparison would require the same area of study under both baselines. What survives across configurations is the geography: Iberian wildfire exposure is a northern and upland phenomenon, and this week's authority alerts and active fires fall inside that geography.

§DModeled risk versus observed fire

Because the FIRMS overlay is an independent, satellite-observed layer, each map carries its own internal check: where a yellow detection lands on or beside a red modeled-risk cell, model and observation agree. In this run that agreement holds at Leciñena and at the Gavarres. This is corroboration of plausibility. It is not evidence of predictive skill, because FIRMS reports fire that is already burning, concurrent with the run, not fire the model called in advance. Skill can only be assessed by tracking outcomes against the flagged cells over the coming weeks, which is the loop this report is designed to feed.

§ERecommended use

1. Treat these surfaces as operational readiness guidance, not as a map of where fire will occur.
2. Commission a prior-dominant structural run over the same domains before siting durable intervention such as fuel management or fuel-break maintenance.
3. Ground-verify the heaviest red clusters first: in Portugal, Águeda–Albergaria and the Castro Daire–Sernancelhe highlands; in Spain, the Guilleries–Montseny massif (the unrealized inland flag) and the Leciñena and Gavarres perimeters.
4. Route the wind-farm finding to grid and renewables operators, given the disproportionate flagged share in Run A.
5. Freeze and date this report at issuance so its scenarios can be evaluated honestly against the season that follows.

§FAssumptions, uncertainty & how to read the figures

The figures rest on three standing assumptions: that the chosen ENSO baseline is the correct reference for the period; the skill of the GFS forecast at horizons approaching 504 hours, which degrades with lead time; and the exposure model's crossing with asset layers, whose denominators are only as current as the underlying cadastres. The triangulation above is drawn from authority statements and press reporting for the run window; it establishes that the modeled surface is plausible and consistent with events on the ground, not that the model is skillful.

The model returns a probability-weighted exposure field, not a point prediction. Percentages are the expected share of each asset within the flagged surface under the stated configuration. They are estimates for planning, not counts of confirmed or expected events, and have not been validated against realized outcomes.

§GSources

• Governo de Portugal / ANEPC, state of alert 3–6 July 2026. portugal.gov.pt
• ANEPC, worsening rural-fire danger, Norte and Centro, from 2 July. portugal.gov.pt
• IPMA, Rural Fire Danger (PIR). ipma.pt
• Euronews, over 1,000 firefighters battle wildfires (03 Jul 2026). euronews.com
• Moncloa, two Aragón fires (Leciñena, La Fueva) ~3,000 ha (03 Jul 2026). moncloa.com
• Ara, fire risk map of Catalonia; La Bisbal d'Empordà / Gavarres fire. ara.cat
• AEMET, forecasted fire-hazard maps. aemet.es

Capybara Center for Risk Analysis · Tecnológico de Monterrey · under the JANUS framework (Feged-Rivadeneira, 2025; archived release doi.org/10.5281/zenodo.21170613). Legacy model run. Probabilistic exposure estimates, not deterministic forecasts of ignition or damage. Triangulation establishes consistency, not predictive skill.