WILDFIRE RISK SITUATION REPORT
WILDFIRE RISK SITUATION REPORT
Colombia | CFSv2 Seasonal Analog | El Niño Phase | March 2026
Integrated Bayesian Wildfire Risk Model — Google Earth Engine | Dual-run Comparative Analysis
| HORIZON | CFSv2 Seasonal | ENSO | El Niño | RES | 500 m |
1. EXECUTIVE SUMMARY
This situation report presents a comparative analysis of two Bayesian wildfire risk model runs for Colombia, both using the CFSv2 seasonal analog forecast under El Niño conditions (years 2019, 2023, 2024). Run 1 covers the 3-month window (March–May) and Run 2 extends to 7 months (March–September), representing the full dry-to-transitional season. The extended run reveals substantially higher expected exposure across all asset classes, with population at risk increasing from 6.9 million (21%) to 8.6 million (26%), and oil palm exposure reaching a critical 65.2%. Three primary high-risk zones are identified: the Caribbean lowlands, the inter-Andean corridors, and the Llanos Orientales foothills.
2. MODEL DESCRIPTION
2.1 How the model works — plain language
The model acts like a risk analyst combining two types of evidence. First, it asks: where have fires actually burned during El Niño years in the satellite record? This creates a memory of fire-prone places. Second, it asks: how dangerous does the weather look for the season ahead? — combining temperature, wind speed, atmospheric dryness, and rainfall deficit into a single hazard score. These two pieces of evidence are fused mathematically so that a place needs both a history of burning AND dangerous weather to score as high risk. The result is a probability map where every 500-metre cell has a score between 0 (safe) and 1 (near-certain fire conditions).
2.2 Processing pipeline
| # | Stage | Data Source | Output |
|---|---|---|---|
| 1 | Historical fire density prior | VIIRS NOAA-20 NRT (2019–2024, El Niño years) | Probability surface of where fires have historically occurred in El Niño conditions |
| 2 | Satellite embedding prior | Google Satellite Embedding V1 (2022) | Landscape similarity score — areas that visually resemble past fire environments score higher |
| 3 | Naïve Bayes calibration | Sampled presence/background from steps 1–2 | Converts raw scores to calibrated probabilities using Laplace-smoothed bin mapping |
| 4 | Fire-weather hazard (p_like) | CFSv2 climatology (1990–2019) | Seasonal fire-weather likelihood from T, wind, humidity, precipitation at 25 km resolution |
| 5 | Bayesian log-odds fusion | Calibrated prior + hazard | Final posterior wildfire probability per pixel (0 = no risk → 1 = maximum risk) |
| 6 | Exposure assessment | GPW, MODIS, GHSL, WDPA, power/grid layers | Posterior-weighted expected exposure per asset class |
2.3 Run configurations
| Parameter | Value |
|---|---|
| Country | Colombia |
| Forecast dataset | CFSv2 (NOAA/CFSV2/FOR6H) — seasonal climatological analog, 1990–2019 |
| ENSO filter | El Niño only — years 2019, 2023, 2024 (ONI ≥ +0.5) |
| Fire baseline points | 111 VIIRS NOAA-20 fire event points |
| Run 1 — target months | March, April, May (3-month dry season core) |
| Run 2 — target months | March – September (7-month full season) |
| Analysis resolution | 500 m (fine) | 2,000 m (display) |
| Prior balance | Balanced — prior weight 0.7, hazard weight 1.4 |
| Embeddings & calibration | Enabled (Google Satellite Embedding V1, Laplace α=1.0) |
3. VARIABLE GLOSSARY
| Variable | Plain Language | Technical Definition |
|---|---|---|
| p_prior | Base fire probability from history. | Kernel density estimate of VIIRS fire detections (2019–2024, El Niño years only), calibrated via naïve Bayesian model with Laplace smoothing (α=1.0). |
| p_like | How dangerous the weather looks this season. | Normalised fire-weather hazard index from CFSv2 climatological mean (T, specific humidity, u/v wind, precip rate), sigmoid-transformed with P8–P92 percentile scaling. |
| p_posterior | Final risk score: history × weather danger combined. | Log-odds fusion: logit(p_prior) + 1.4 × logit(p_like), passed through sigmoid. Output range [0,1]. |
| ENSO filter | Only uses El Niño years — Colombia is drier and more fire-prone during El Niño. | VIIRS archive stratified by NOAA ONI threshold ≥+0.5. Selects 2019, 2023, 2024. |
| CFSv2 analog | Average of past climate data for the same months ahead. | NCEP CFS v2 (NOAA/CFSV2/FOR6H). One image per year/month (1990–2019), composited at 25 km. |
| Expected exposure | How much of each asset is expected to be affected, weighted by fire probability. | E[exposed] = ∫ p_posterior(x) · exposure(x) dA. Pixel-area-weighted reduceResolution aggregation. |
| Seasonal window | How many months ahead the model looks. | Run 1: N=3 (Mar–May). Run 2: N=7 (Mar–Sep). CFSv2 mean across 30 years per month. |
| N/A (roads/RE/oil) | These layers returned no values in this run. | Artefact from raster resolution and sparse point density at pointSampleScale=250m. |
4. MODEL OUTPUT MAPS
4.1 Run 1 — 3-Month Seasonal Window (March–May)
ENSO: El Niño (2019, 2023, 2024) | CFSv2 months 3, 4, 5 | fineScale=500 m
RUN 1 · CFSv2 Mar–May · El Niño · 21.0% Population at Risk (6.9M)
Run 1 — CFSv2 3-Month Seasonal Analog (Mar–May) | El Niño | Colombia | fineScale=500m
Key observations: Risk is concentrated in the Caribbean coastal lowlands (La Guajira, Cesar, Barranquilla region) and the northeastern inter-Andean corridor through Santander and Norte de Santander toward the Venezuela border. A secondary elevated band is visible in the Magdalena valley (Huila/Tolima). The eastern Llanos show moderate risk. Southern Andean departments and Amazonia remain low risk.
4.2 Run 2 — 7-Month Seasonal Window (March–September)
ENSO: El Niño (2019, 2023, 2024) | CFSv2 months 3–9 | fineScale=500 m
RUN 2 · CFSv2 Mar–Sep · El Niño · 26.0% Population at Risk (8.6M)
Run 2 — CFSv2 7-Month Seasonal Analog (Mar–Sep) | El Niño | Colombia | fineScale=500m
Key observations: The 7-month window shows substantial spatial expansion. The Magdalena valley band widens and intensifies. Andean flanks from Cundinamarca through Huila and Caquetá enter elevated risk. The eastern Llanos (Meta, Casanare) show deeper penetration. Oil palm exposure jumps to 65.2%. Electric grid exposure reaches 32.9% (291 km).
5. COMPARATIVE EXPOSURE ANALYSIS
The table below compares posterior-weighted expected exposure between the two runs. All increases are attributable to the extended seasonal window drawing additional months of fire-weather hazard into the climatological mean.
| Variable | Run 1 — 3-Month (Mar–May) | Run 2 — 7-Month (Mar–Sep) | Δ Change |
|---|---|---|---|
| Population at risk | 21.0% · 6,911,288 | 26.0% · 8,564,969 | +5.0pp · +1.65M |
| Cropland area | 18.0% · 82.96 ha | 34.1% · 157.22 ha | +16.1pp · +89% |
| Urban area | 29.6% · 74.65 ha | 36.0% · 90.96 ha | +6.4pp · +22% |
| Oil palm area | Computing… | 65.2% · 61.29 ha | ↑ HIGH |
| Protected areas | 7.7% · 925 ha | 9.3% · 1,114 ha | +1.6pp · +20% |
| Built-up footprint | 19.0% · 13,153 ha | 19.6% · 13,547 ha | +0.6pp · +3% |
| Electric grid | Computing… | 32.9% · 291.23 km | ↑ HIGH |
| Roads / Oil & Gas / RE | N/A | N/A | — |
Key findings:
- Population: +1.65 million additional people in risk zones (26% vs 21%). Progressive drying of Andean flanks and Llanos communities as season extends into June–September.
- Cropland: Exposure nearly doubles (+89%), driven by Llanos agricultural frontier and Magdalena valley farmland.
- Oil palm (CRITICAL): 65.2% in Run 2 — highest exposure of any variable. Palm zones in Meta, Casanare, and Magdalena Medio acutely vulnerable to extended dry seasons.
- Electric grid: 291 km of transmission/distribution lines at risk (33% in Run 2). Grid disruption would cascade to water pumping and emergency response.
- Protected areas: 9.3% of mapped protected area (1,114 ha) at risk in Run 2, up from 7.7%. Includes páramo zones in Boyacá/Santander.
6. SPATIAL FINDINGS BY ZONE
6.1 Caribbean lowlands — CRITICAL (both runs)
- Departments: La Guajira, Cesar, Magdalena, Córdoba, northern Bolívar
- Deep red signal in both runs. Strong NE trade winds + persistent humidity deficit + high historical VIIRS fire density = highest combined posterior in the country.
- Consistent with UNGRD 2024 emergency declarations for La Guajira and Cesar.
6.2 Inter-Andean corridors — HIGH (expanded in Run 2)
- Departments: Santander, Norte de Santander, Boyacá, Cundinamarca, Huila, Tolima
- Clear north–south risk band in rain-shadow valleys between cordilleras.
- Run 2 shows significant southward expansion into Huila and Caquetá.
6.3 Llanos Orientales — HIGH in Run 2 (MOD in Run 1)
- Departments: Meta, Casanare, Vichada, Arauca
- Savanna biome with strong seasonal fire cycle. Run 2 captures the mid-year second dry season (June–August).
- Oil palm concentration in Meta/Casanare directly explains the 65% oil palm exposure in Run 2.
6.4 Pacific coast and Amazonia — LOW (both runs)
- Departments: Chocó, Nariño, Putumayo, Amazonas, Vaupés
- High precipitation and low humidity deficit suppress CFSv2 hazard. Few historical fire events provide a weak prior.
7. FORECASTING AND DECISION OBJECTS (FDOs)
FDOs follow the JANUS analytic method, synthesising open-source intelligence with direct model output. Two registers: plain language for operational users and expert definition for technical review.
| FDO 01 El Niño amplifies Colombian dry-season fire risk by suppressing rainfall east of the Andes | |
| Source | NOAA ONI; ReliefWeb ACAPS 20 Feb 2024; Scientific American 22 Feb 2024; CNN 25 Jan 2024 |
| Simple | During El Niño, Colombia gets far less rain than normal — especially in the east. This dries out forests and grasslands, making fires much more likely and much harder to control. |
| Expert | El Niño positive phase (ONI ≥+0.5) suppresses ITCZ precipitation over the Colombian Andes and Llanos via anomalous descending motion. Results in soil moisture deficit, elevated VPD, and increased fine-fuel flammability across the Orinoquía and Caribbean biomes. |
| Indicator | ENSO filter active (2019, 2023, 2024). 111 VIIRS fire event points. CFSv2 seasonal mean confirms elevated T and low specific humidity. Both runs show high posterior risk in ENSO-consistent hotspot zones. |
| Implication | Hotspot zones should trigger pre-positioned suppression resources and community early warning activation at the onset of every confirmed El Niño event, from January through September. |
| FDO 02 Extended seasonal window (7 months) adds 1.65 million people to the wildfire risk population | |
| Source | Model output — dual run comparison; CIESIN GPWv4; UNGRD 2024 |
| Simple | When the model looks further into the future, 1.65 million more Colombians are estimated to live in fire-risk areas. The number jumps from 6.9 million to 8.6 million people. |
| Expert | Posterior-weighted population integral increases from 6,911,288 (Run 1, 21.0%) to 8,564,969 (Run 2, 26.0%). The marginal 1,653,681 persons enter risk as CFSv2 July–September months introduce the second Colombian dry season in Meta, Casanare, and southern Andean departments. |
| Indicator | +1.65M, +5.0 percentage points. Spatial expansion visible in Llanos departments and Andean southern flank in Run 2 map. |
| Implication | Use 8.6M for worst-case contingency planning. The 6.9M figure remains appropriate for immediate Q1 operational response. |
| FDO 03 Oil palm exposure reaches 65% in the 7-month run — critical agricultural risk | |
| Source | Model output — Run 2; BIOPAMA GlobalOilPalm V1; FAO Colombia land use data |
| Simple | When the model extends to 7 months, nearly two thirds of Colombia's mapped oil palm area falls within wildfire risk zones. Palm is Colombia's leading agricultural export, concentrated in Meta, Casanare, and Magdalena Medio. |
| Expert | 65.2% × 93.99 ha total = 61.29 ha at risk. BIOPAMA GlobalOilPalm V1 classes 1 and 2 (industrial and smallholder). Spatial correlation with Llanos Orientales high-risk zone confirmed. |
| Indicator | Pixel-area-weighted posterior integral at 250m scale. Run 2 Llanos expansion directly accounts for the increase. |
| Implication | Fedepalma should be notified when Run 2-level risk is forecast. Firebreak maintenance in Meta and Casanare should be mandated before June onset of the second dry season. |
| FDO 04 291 km of electric grid infrastructure at wildfire risk in the full 7-month model | |
| Source | Model output — Run 2; SAT-IO predictive power system; Facebook GMV grid layers |
| Simple | Almost 300 km of electricity lines pass through high-risk wildfire zones. Damage could cut power, stop water pumps, and disable emergency communications. |
| Expert | 32.89% × 885.36 km total = 291.23 km at risk. Grid layer from SAT-IO/predictive-global-power-system and Facebook GMV. Spatial correlation with Llanos and inter-Andean corridor hotspots confirmed. |
| Indicator | 291.23 km at risk in Run 2. Grid risk tracks the Magdalena valley corridor and Meta oil-field transmission lines. |
| Implication | XM and UPME should integrate this model into annual grid resilience planning. Fire-weather alerts should trigger pre-emptive line inspections within the 291 km risk corridor. |
| FDO 05 Páramo wetlands — supplying 85% of Colombia's drinking water — show increasing fire exposure | |
| Source | Scientific American 22 Feb 2024; Aguilar Garavito (Pontifical Xavierian Univ.); IDEAM Jan 2024; model output |
| Simple | Colombia's high-altitude wetlands are now burning far more often than historically. They supply most of the country's drinking water. Risk grows in the 7-month window. |
| Expert | Pre-industrial páramo fire return interval: 100–1,000 years. Current observed: 2–10 years. Protected area exposure increases from 925 ha (7.7%) to 1,114 ha (9.3%). WDPA analysis identifies Boyacá-Santander páramo complexes within the risk perimeter. |
| Indicator | +189 ha between runs (+20%). Spatial expansion in Run 2 aligns with the southern Andean risk corridor. |
| Implication | Any fire alert within 5 km of a gazetted páramo boundary should trigger immediate aerial suppression. Bogotá and Medellín utilities should link model outputs to reservoir contingency protocols. |
| FDO 06 Climate change is intensifying El Niño fire impacts beyond what CFSv2 climatology captures | |
| Source | Scientific American 22 Feb 2024; IDEAM Jan 2024; World Weather Attribution 2024; Axios 29 Jan 2024 |
| Simple | Climate change is making El Niño events hotter and more dangerous than the historical average used in this model. January 2024 was Colombia's hottest month in 30 years. The model likely underestimates future risk. |
| Expert | World Weather Attribution (2024): human-induced warming is primary catalyst for 2023–24 Amazon drought severity. IDEAM confirmed January 2024 as hottest since 1994 (+5–10°C anomaly). CFSv2 1990–2019 baseline does not incorporate post-2015 accelerated warming. Posterior risk likely underestimated by 10–25%. |
| Indicator | CFSv2 1990–2019 baseline is ~0.8–1.2°C cooler than present-day El Niño conditions. Hazard index normalisation may be systematically shifted low for post-2020 events. |
| Implication | Supplement with the CMIP6 long-term scenario layer (SSP3-7.0 or SSP5-8.5). The CFSv2 run should be treated as a conservative lower bound on actual current-season risk. |
8. MODEL LIMITATIONS & CAVEATS
- CFSv2 is a climatological archive (1990–2019), not a real-time forecast. Use GFS short-term mode for operational daily risk.
- 25 km native resolution means local topographic features are not resolved. Interpret at department or regional scale only.
- Human ignition probability is absent — 95% of Colombian fires are human-started (MADS 2024), a structural model gap.
- Roads, oil & gas wells, and RE facilities show N/A due to raster resolution limitations. Adjusting roadRasterScale and pointSampleScale will resolve these.
- ENSO classification uses a hardcoded 2019–2025 lookup table. Update annually with NOAA CPC ONI classifications.
- WDPA boundaries may not reflect recent updates. Validate against current RUNAP registry.
- Oil palm layer (BIOPAMA GlobalOilPalm V1) was published in 2021 — recent expansion in Meta and Casanare may be underestimated.
Document generated: March 2026 | Model: Integrated Bayesian Wildfire Risk Model v2 (GEE) | ENSO: El Niño (2019, 2023, 2024) | Forecast: CFSv2 Seasonal Analog | Run 1: Mar–May | Run 2: Mar–Sep | Classification: UNCLASSIFIED / FOR PLANNING USE
Note: Only a member of this blog may post a comment.