WILDFIRE RISK SITUATION REPORT

 

Mexico (Northwest Corridor) | WeatherNext 2 Ensemble | El Niño Phase | 15-Day Horizon | June 2026

Integrated Bayesian Wildfire Risk Model — Google Earth Engine | First WeatherNext 2 run for Mexico

HORIZON	WeatherNext 2 (≤360 h, 15-day ensemble)	ENSO	El Niño (2019, 2023, 2024)	RES	3 km fine / 20 km display (big-country profile)

---

1. EXECUTIVE SUMMARY

This report documents the first run of the Integrated Bayesian Wildfire Risk Model for Mexico using the WeatherNext 2 (WN2) ML ensemble forecast. Under an El Niño-filtered prior, the 15-day forecast concentrates posterior wildfire risk in a single northwest corridor stretching from Baja California across Sonora into northern Chihuahua. The signal is geographically narrow but structurally consequential: it captures the bulk of Mexico's mapped solar capacity (60.96% of plotted sites), a fifth of the country's protected-area footprint inside the model AOI (21.45%), and a fifth of its built-up land (21.06%). Population exposure is comparatively modest (5.19%, ~167k people) because the corridor crosses the country's least-densely-populated belt. The full population figure is anomalously low and is treated as a known artefact of the big-country profile (see §8).

---

2. MODEL DESCRIPTION

2.1 How the model works — plain language

The model behaves like a two-source risk analyst. Source 1: where have wildfires actually burned in El Niño years according to the satellite record? That answer becomes the prior. Source 2: how dangerous does the weather look for the next 15 days according to a 64-member ML ensemble forecast? That becomes the hazard. The two are fused mathematically so a pixel only scores high when both evidence types agree. A separate exposure step then weighs every map pixel by the posterior probability and counts how much population, cropland, infrastructure and protected land sits in the risk zone.

2.2 Processing pipeline

#	Stage	Data source	Output
1	Historical fire density prior	VIIRS NOAA-20 NRT (2019–2025, ENSO-filtered to El Niño years)	Where fires have actually occurred in El Niño
2	Satellite embedding prior	Google Satellite Embedding V1 (2022)	Landscape similarity to past fire pixels
3	Naïve Bayes calibration	Sampled presence/background pixels	Calibrated prior probability via Laplace-smoothed binning
4	Fire-weather hazard (p_like)	WeatherNext 2 ensemble — 2m_temperature, 1000_specific_humidity, 10m_u/v_wind, total_precipitation_6hr; latest init, 64 members averaged across 0–360h leads	Fire-weather likelihood from a state-of-the-art ML ensemble
5	Bayesian log-odds fusion	Calibrated prior + hazard	Posterior wildfire probability
6	Exposure assessment	GPWv4.11, MODIS MCD12Q1, GHSL Built, BIOPAMA Oil Palm, WDPA, SAT-IO/Facebook grid, WRI GPPD power plants	Posterior-weighted expected exposure by asset class

2.3 Run configuration

Parameter	Value
Country	Mexico (FAO GAUL ADM0)
Forecast dataset	WeatherNext 2 — projects/gcp-public-data-weathernext/assets/weathernext_2_0_0
Ensemble strategy	All 64 members averaged at native 25 km, then upsampled for display
Lead window	15 days (≤360 forecast hours)
ENSO filter	El Niño only — 2019, 2023, 2024
Fire baseline points (priors)	115 VIIRS NOAA-20 fire-event sample
Country size profile	big_country → fineScale = 3 km, displayScale = 20 km
Prior balance	Balanced — prior weight 0.7, hazard weight 1.4
Embeddings & calibration	Disabled by big_country automatic flag

---

3. VARIABLE GLOSSARY

Variable	Plain language	Technical definition
p_prior	Base fire probability from history	VIIRS fire-density kernel (50 km, El Niño years) blended in logit space
p_like	How dangerous the next 15 days look	Hazard index from WN2 ensemble mean of T, SH₁₀₀₀, wind, precip; percentile-stretched and sigmoid-sharpened
p_posterior	Final risk score	sigmoid( logit(p_prior) + 1.4 · logit(p_like) )
WN2 ensemble	Average of 64 forecast members from the latest init	Initialised every 6h, 6-hour lead-time steps, 0.25° resolution, max 15-day horizon
Expected exposure	Asset area/count weighted by posterior probability	∫ p_posterior(x) · exposure(x) dA
Big-country profile	Coarser resolution to keep large AOIs tractable	fineScale=3 km, displayScale=20 km, embedding+calibration auto-disabled

---

4. MODEL OUTPUT MAP


Wildfire posterior risk — Mexico | WeatherNext 2 ensemble (≤360 h, 15-day window) | ENSO: El Niño (2019, 2023, 2024) | fineScale=3 km, displayScale=20 km

Key observations:

• The signal forms a single contiguous northwest corridor: Mexicali region of Baja California Norte → all of Sonora → northern and central Chihuahua → northwestern Durango.
• Highest posterior cells sit on the Pacific-facing slope of the Sierra Madre Occidental and the Sonoran piedmont.
• A secondary, weaker plume reaches into Sinaloa and edges of Coahuila.
• Central, southern, and Yucatán Mexico register near-zero posterior in this window. The WN2 ensemble is showing a clearly anisotropic continental pattern.
• The geographic envelope matches established El Niño-phase fire ecology for northwest Mexico (Sonoran/Chihuahuan deserts and the rain-shadow side of the Sierra Madre).

---

5. EXPOSURE ANALYSIS

Posterior-weighted expected exposure inside the Mexico AOI for the 15-day WN2 window.

Variable	% of total	Expected (absolute)	Total (denominator)
Population (expected exposed)	5.19%	166,694 ppl	3,210,887 ppl ⚠
Cropland area	8.67%	18,757.95 ha	216,252.02 ha
Urban area	6.84%	906.37 ha	13,243.38 ha
Oil palm area	0.17%	0.03 ha	16.66 ha (negligible in Mexico)
Protected areas	21.45%	94,603.49 ha	441,122.01 ha
Built-up footprint	21.06%	627,732.85 ha	2,980,648.54 ha
Electric grid length	N/A	—	(raster-resolution artefact)
Roads length	N/A	—	(raster-resolution artefact)
Oil & gas sites	N/A	—	(point-sampling-scale artefact)
Wind farms	4.93%	1 site	20 sites
Solar farms	60.96%	2 sites	3 sites

⚠ The population denominator (3.21 M) is implausibly low for Mexico (~130 M nationally). Treated as a known artefact of the big-country profile, not as evidence about Mexico's true population. See §8 and FDO 06.

---

6. FORECASTING AND DECISION OBJECTS (FDOs)

FDOs follow the JANUS analytic method (Feged, 2025, https://doi.org/10.21203/rs.3.rs-7580516/v1): every assertion below links to ≥1 evidence span — here, a specific dataset and computation in the model run. Two registers are kept: Simple (plain language for operational users) and Expert (technical detail for review). Each FDO ends with an explicit Implication (decision-relevant action) and Limits (epistemic boundary).

FDO 01 — Wildfire risk is concentrated in a single northwest corridor in the 15-day WN2 window

Field	Content
Source / trace	WN2 posterior layer (final_threat), 64-member ensemble mean over 0–360h leads from latest init; ENSO filter active (El Niño 2019/23/24); VIIRS fire-density prior over the same years
Simple	The next 15 days concentrate fire-weather risk in northwest Mexico — Baja California, Sonora, and northern Chihuahua. The rest of the country is below detection threshold in this window.
Expert	Posterior support spans roughly 28–32°N, 110–106°W, aligning with the Sonoran/Chihuahuan dry belt and the rain-shadow flank of the Sierra Madre Occidental. The hazard component p_like dominates outside historical VIIRS hotspots, suggesting the WN2 ensemble is forecasting a meso-α drying signal rather than reactivating a uniformly weighted prior.
Indicator	Visible posterior > 0.5 only in the NW corridor; remaining states near zero.
Implication	CONAFOR and SEMARNAT regional offices in Baja California, Sonora, Chihuahua and Sinaloa should treat the 15-day window as elevated. Other states are NOT cleared — they are below the model's current detection horizon.
Limits	A 3 km fine / 20 km display run resolves macro-corridors only. Substate-scale variation is not retrievable from this run.

FDO 02 — Two-thirds of Mexico's mapped utility-scale solar capacity is inside the wildfire corridor

Field	Content
Source / trace	WRI GPPD power_plants filtered to fuel1 IN ('Solar','SOLAR','solar','Solar PV','SolarPV','PV'); posterior sampled at point locations at pointSampleScale = 250 m
Simple	Of the 3 utility-scale solar farms the WRI database has mapped in Mexico, the model places 2 inside the 15-day fire-weather corridor. That is two-thirds of the country's mapped solar fleet.
Expert	Expected posterior-weighted count = 2 of 3 (60.96%). Spatial correlation with the Sonora corridor is consistent with the location of large operating PV plants (Puerto Libertad, Hermosillo cluster).
Indicator	Solar farms — expected exposed: 2 of 3 sites (60.96%) — highest single-variable exposure in the run.
Implication	Asset operators with PV in Sonora should run their fire-defence and panel-cleaning protocols ahead of the window. SENER and CRE should consider whether the WRI database is materially incomplete — this denominator may understate national solar capacity, in which case the 60.96% figure represents an upper bound on the mapped fleet, not on the total fleet.
Limits	WRI GPPD is incomplete for newer Mexican solar buildout (post-2022). Treat 60.96% as "exposure of mapped sites", not "exposure of national solar capacity".

FDO 03 — One-fifth of Mexico's protected-area footprint is inside the 15-day risk perimeter

Field	Content
Source / trace	WCMC WDPA current/polygons, rasterised at protectedAreaScale = 2.5 km (big-country profile); posterior-weighted area integration
Simple	The model places 21.45% of mapped protected-area hectares inside the risk perimeter. Almost 95,000 ha.
Expert	94,603.49 ha at risk of 441,122.01 ha total inside the Mexico AOI computation. WDPA polygons in the Sonoran corridor include El Pinacate y Gran Desierto de Altar Biosphere Reserve (UNESCO World Heritage), Alto Golfo de California, Sierra de Álamos / Río Cuchujaqui, and several Áreas de Protección de Flora y Fauna.
Indicator	Protected areas — expected exposed: 21.45% · 94,603 ha.
Implication	CONANP should pre-position fire-suppression resources at the El Pinacate, Alto Golfo, and Sierra de Álamos perimeter offices. Biosphere reserve buffer zones should be specifically advised.
Limits	WDPA "polygons" include both terrestrial and marine protected areas; the marine fraction is irrelevant to wildfire risk but is not separated in this run. The 441,122 ha denominator is therefore an underestimate of how much protected land would be the more correct base.

FDO 04 — 21% of mapped built-up land is in the wildfire corridor

Field	Content
Source / trace	JRC GHSL BUILT_LDSMT_GLOBE_V1 (built > 0), rasterised at builtupScale = 2.5 km; posterior-weighted area integration
Simple	About 627,000 hectares of Mexico's mapped built-up land — roads, buildings, paved surface — sit inside the 15-day fire-weather corridor.
Expert	627,732.85 ha of 2,980,648.54 ha (21.06%). The built-up "footprint" in the corridor is the wildland–urban interface around Hermosillo, Ciudad Obregón, Ciudad Juárez and the Mexicali-Calexico borderland, plus the smaller northern Sonoran towns.
Indicator	Built-up footprint — expected exposed: 21.06% · 627,733 ha.
Implication	Civil protection coordination should focus on the wildland–urban interface zones around Hermosillo, Cd. Obregón, and Cd. Juárez — these are the population centres where built-up footprint coincides with corridor risk.
Limits	GHSL P2016 is dated. Recent peri-urban expansion (post-2016) is not captured; the built-up denominator is likely a lower bound.

FDO 05 — WeatherNext 2 extends operational forecast horizon to 15 days at ensemble strength, vs. ~3 weeks deterministic with GFS

Field	Content
Source / trace	Model run with forecastHorizon = short_term_wn2, wn2LeadDays = 15; WN2 dataset metadata; comparison with prior runs at forecastHorizon = short_term (GFS)
Simple	This is the first run of the wildfire model using Google DeepMind's WeatherNext 2 ensemble. It looks 15 days ahead, averages 64 possible weather futures, and at 0.25° resolution. Earlier runs used GFS, which is deterministic and noisier at long lead times.
Expert	WN2 provides probabilistically-calibrated multi-day forecasts via a Functional Network Generative model. For wildfire pre-positioning, the ensemble-mean signal at 7–15 days is the operationally useful new capability — GFS deterministic skill collapses well before 360h.
Indicator	1,792 ensemble images successfully retrieved and averaged from the latest init; band coverage confirmed (2m_temperature, 1000_specific_humidity, 10m_u/v_wind, total_precipitation_6hr).
Implication	Add WN2 short-term-ML as a routine companion run to the GFS short-term run. Discrepancies between the two are informative — they bound the "ensemble of forecast methods" uncertainty.
Limits	WN2 is experimental and access-gated (Google data-request form). The ensemble mean smooths extremes; for tail-risk analysis a future iteration should report ensemble spread / 90th-percentile fields, not only the mean.

FDO 06 — The 3.2 M population denominator is an artefact of the big-country profile, not evidence about Mexico

Field	Content
Source / trace	Population layer CIESIN/GPWv411/GPW_Population_Count reduced at populationScale = 5 km (auto-set by big_country profile); reduceRegion bestEffort = true
Simple	The model reports 3.21 M total population for Mexico. The actual figure is ~130 M. The reported number is what the model could "see" at 5 km resampled resolution with best-effort aggregation — not what Mexico actually contains.
Expert	At big_country profile, populationScale is forced to ≥5 km and bestEffort: true is set on the reduceRegion. The result undercounts dense pixels via the bestEffort projection. The ratio (5.19% expected of total) remains internally consistent — the 166,694 expected vs 3.21 M total is a valid relative metric within this run. Cross-run absolute comparisons are not.
Indicator	Population total in panel: 3,210,887 ppl (compare: World Bank 2024 Mexico population ≈ 130,000,000).
Implication	Do not quote "166,694 people at risk" as an absolute Mexican figure. Report only the fraction (5.19% of GHSL-mapped-and-resampled population in the AOI) and note the artefact. For absolute population-at-risk numbers, re-run at country_size_profile = standard with a focused AOI buffer over Sonora-Chihuahua.
Limits	Identified inside this report; not yet patched in the model. Fix is straightforward: drop bestEffort and lower populationScale to 1 km when a sub-national AOI is used.

---

7. SPATIAL FINDINGS BY ZONE

7.1 Sonora–Chihuahua corridor — CRITICAL

• States: Baja California Norte, Sonora, Chihuahua, Durango (north), Sinaloa (north).
• Drivers: persistent humidity deficit on the Pacific-facing flank, sustained westerlies, and historical El Niño dryness anomalies.
• Anchors: El Pinacate y Gran Desierto de Altar, Sierra de Álamos, Sierra Madre Occidental piedmont.

7.2 Central plateau / Bajío — LOW

• States: Aguascalientes, Guanajuato, Querétaro, Hidalgo, Edomex.
• WN2 signal absent in this window. Consistent with El Niño pattern (drying is north-biased in Mexico, opposite to its Pacific-South-America signal).

7.3 Pacific coast (south) / Yucatán / Chiapas — LOW

• High mid-tropospheric moisture from incipient North-American monsoon onset suppresses fire-weather. WN2 ensemble agrees with climatological expectation.

---

8. MODEL LIMITATIONS & CAVEATS

• Population artefact (FDO 06). Absolute population-at-risk number is not usable. Ratio is fine.
• big_country profile auto-disables embeddings and calibration. The prior collapses to VIIRS density only, which is more spatially smooth than the embeddings-augmented variant. Re-running on a sub-national AOI would restore both stages.
• N/A rows for electric grid, roads, oil & gas sites. These are not "zero exposure" findings — they are raster-resolution artefacts. At pointSampleScale = 250 m against a 20 km display, sparse point features fall between samples. Re-running at standard profile fixes this.
• WN2 is experimental. Access is gated by the WeatherNext Data Request form. The ensemble mean is reported here; for operational decisions, ensemble spread should also be inspected.
• ENSO classification is hardcoded in a 2019–2025 lookup table inside the model. Update annually with NOAA CPC ONI.
• WDPA + GHSL vintages (2016/2024 respectively) mean recent infrastructure and protected-area changes may not appear.

---

9. AI-USE DISCLOSURE & REFLEXIVE MEMO (JANUS framework)

This is a JANUS Section-5 disclosure (Feged, 2025).

LLM assistance. A large language model assisted with: (a) structuring the situation-report sections; (b) drafting the plain-language Simple text of each FDO; (c) suggesting cross-references between observable indicators and likely implications; (d) clarifying phrasing throughout. The LLM did not: generate model output, fabricate exposure numbers, invent dataset names, infer ENSO years, or write any Source / trace, Indicator, or Limits line without a corresponding artefact in the underlying GEE run.

Provenance. All numerical exposure figures are read directly from the model's bottom-left exposure panel as printed in the source image; the run configuration is read from the panel header. The model code itself is versioned at /PROD/src/earth-engine/heatwaves-forest-fires/interactive-wildfire-risk-model.js and includes explicit WN2 diagnostic prints ([WN2 DIAG]) confirming image counts and band availability for this run.

Reflexive memo (analyst positionality). The analyst is an outsider to operational Mexican wildfire response and has not consulted CONAFOR, SEMARNAT or CONANP for ground-truth corroboration of the northwest-corridor signal at the time of writing. The FDO implications should be treated as analytic suggestions, not as operational orders. The post is published as a Situation Report, not as a decision.

Claim–evidence contract. No assertion in §§5–7 is published without a trace to either a model layer (with its EE dataset ID) or to a clearly labelled visual feature of the output map. The N/A rows are explicitly not used to make exposure claims.

Adversarial pass. Three plausible alternative explanations for the signal were considered and rejected:

1. Prior-only artefact — rejected because removing the prior leaves the same corridor in the hazard layer.
2. Coarse-resolution smearing — partially accepted (see §8): the corridor is real, but its sharp edges are likely smoother in reality.
3. Cherry-picked ENSO years — the El Niño years 2019, 2023, 2024 are not arbitrary; they are the post-2018 set with ONI ≥ +0.5. Re-running with all_years confirms the corridor is still the dominant feature, though less pronounced.

---

Document generated: 2 June 2026 | Model: Integrated Bayesian Wildfire Risk Model (GEE, v3 w/ WN2) | Forecast: WeatherNext 2 ensemble, 15-day, latest init | ENSO: El Niño (2019, 2023, 2024) | Country profile: big_country | Classification: UNCLASSIFIED / FOR PLANNING USE

Labels: Early Warning System, Situation Report, Mexico, WeatherNext 2