Colombia | GFS Short-Term Forecast | Bayesian Multi-Hazard Model | June 2026
FLOOD RISK SITUATION REPORT
Colombia | GFS Short-Term Forecast | Bayesian Multi-Hazard Model | June 2026
Integrated Multi-Hazard Model (Fable-tuned, Bayesian flood class) — Google Earth Engine | 7-day Forecast Horizon | Flood-density + Satellite-Embedding Priors + Naïve Calibration + Bayesian Log-Odds Fusion | Includes media-triangulation FDOs (UNGRD / IDEAM / El Tiempo / Semana / regional sources)
| HORIZON | GFS short-term (≤168 h) | ENSO BASELINE | La Niña (2020, 2021, 2022, 2025) — see FDO 11 |
| FUSION | Bayesian (balanced) | RES | 500 m fine / 1 km exposure |
1. Executive summary
The Bayesian flood model places the dominant 7-day flood-threat surface for Colombia across six concurrent corridors: La Guajira and the Caribbean coast (Atlántico, Bolívar, Magdalena, Sucre, Córdoba); the Andean cordillera (Antioquia, Caldas, Risaralda, Quindío, Valle, Cauca, Nariño, Cundinamarca, Boyacá); the eastern Llanos (Arauca, Casanare, Meta, eastward into Vichada); the Pacific corridor (Chocó, coastal Valle and Cauca); and parts of Amazonia (Putumayo, Caquetá, Amazonas). Population exposure is the headline finding: 39.40 percent of mapped population — approximately 12.96 million of 32.88 million people in the model's national base — falls inside the posterior exposure surface.
The corridor structure reproduces, with one synthetic surface, what UNGRD has been reporting in fragmented bulletins through the first half of 2026: more than 600 rain-related emergencies across 328 municipalities in 24 departments, affecting roughly 155,000 families, with mass movements and floods as the dominant event types. The model's three highest-density zones — the central Andes, the Llanos, and La Guajira — align with named events on the ground: Arauquita's public-calamity decree, the Río Meta overflow at Cabuyaro, the Río Metica disaster at Puerto López, and the Riohacha emergency activation.
Two findings warrant operator attention specifically. Oil palm exposure at 60.94 percent is extraordinarily high — the model places roughly three of every five mapped hectares of Colombian palm cultivation inside the posterior surface, concentrated along the Magdalena Medio and the eastern Llanos plantations. Electric grid length at 49.80 percent and urban area at 50.20 percent suggest that the half-of-everything pattern is not isolated: this run is reading a Colombia where wet-season dynamics are activating across almost every populated lowland.
Area-class absolute hectare totals carry the documented defect from prior runs (FDO 12): percentages remain interpretable as ratios of expected exposure to model-mapped totals, but the absolute hectare counts should not be quoted in public communications until the area-aggregation chain is patched. Population, infrastructure-length, and point-asset findings are unaffected.
2. Exposure analysis
Posterior-weighted expected exposure inside the Colombia AOI for the 7-day GFS window under the La Niña-baselined Bayesian fusion. ⚠ marks rows where absolute totals are affected by the area-aggregation defect (FDO 12); percentages remain valid.
| Variable | % of total | Expected (absolute) | Total (denominator) |
|---|---|---|---|
| Population (expected exposed) | 39.40% | 12,957,478 ppl | 32,884,748 ppl |
| Cropland area | 47.30% | 351,003 ha ⚠ | 742,123 ha ⚠ |
| Urban area | 50.20% | 210,793 ha ⚠ | 419,870 ha ⚠ |
| Oil palm area | 60.94% | 236,882 ha ⚠ | 388,718 ha ⚠ |
| Protected areas | 29.25% | 5,775,479 ha ⚠ | 19,741,977 ha ⚠ |
| Built-up footprint | 36.77% | 41,625,432 ha ⚠ | 113,214,464 ha ⚠ |
| Electric grid length | 49.80% | 27,434 km ⚠ | 55,090 km |
| Roads length | 44.83% | 137,559 km | 306,839 km |
| Oil & gas sites | N/A — sparse-point artefact | ||
| Wind farms | N/A — sparse-point artefact | ||
| Solar farms | N/A — sparse-point artefact | ||
Robust headlines for public communication: Population 39.40 % (≈ 12.96 M people). Electric grid 49.80 %. Urban 50.20 %. Oil palm 60.94 %. These are the percentages volunteers and partners can quote directly. Absolute hectare values are flagged for the area-aggregation patch (FDO 12).
3. Spatial findings by zone
| La Guajira & central Caribbean coast — CRITICAL | Continuous posterior across La Guajira peninsula (Riohacha → Maicao → Manaure → Uribia) and the Atlántico–Bolívar–Magdalena–Sucre–Córdoba arc. Hazard-driven and prior-driven both: regional flash-flood archive in the rescaled prior, plus Caribbean rain through the first-rainy-season tail. |
| Andean cordillera — CRITICAL | Continuous from Antioquia through the Eje Cafetero (Caldas, Risaralda, Quindío) into Valle, Cauca, Nariño, and across the Sabana de Bogotá and Boyacá. Orographic intercept × per-basin precip deviation × steep terrain. Coffee belt, banana corridor, and the Bogotá–Medellín–Cali population spine. |
| Eastern Llanos — CRITICAL | Continuous across Arauca, Casanare, Meta, eastward into Vichada. Forecast-driven alluvial-plain hazard. Strongest model-to-ground convergence in the country sits here (see FDO 08). |
| Pacific corridor — HIGH | Chocó (Quibdó, Istmina, Tadó), coastal Valle (Buenaventura), Cauca, Nariño. The wettest place in South America reads steadily as moderate-to-high posterior across the window. |
| Amazonia — HIGH | Putumayo, Caquetá, Amazonas. Lower population density softens public-communication priority, but indigenous-territory access routes are exposed; the Mocoa landslide history reinforces the prior. |
| Magdalena Medio palm belt — OPERATOR-SPECIFIC | The 60.94 % oil-palm exposure concentrates here (Santander–Cesar–Bolívar corridor) and in the Llanos plantation belt. Specific Fedepalma and large-producer attention warranted. |
4. Forecasting and Decision Objects (FDOs) — JANUS method
FDOs follow the JANUS analytic method (Feged 2025, DOI 10.21203/rs.3.rs-7580516/v1). Each FDO links a claim to a dataset, a model layer, or a verified external source. Model FDOs come first; media-triangulation FDOs come second; the adversarial finding closes the section.
Model-derived FDOs
FDO 01 — Six concurrent flood corridors, country-wide exposure
| Source / trace | Posterior risk surface (final_threat) at fineScale = 500 m, visible on the output as a continuous purple field across the named departments. |
| Simple | The model places significant flood risk in six different regions of Colombia at once: La Guajira, the Caribbean coast, the Andes, the Llanos, the Pacific, and Amazonia. |
| Expert | Six spatially distinct corridors each generated by a different mechanism. La Guajira and Caribbean coast: prior-reinforced (regional flood archive) plus active Caribbean rain. Andean cordillera: orographic intercept × per-basin precip deviation × steep slope. Llanos: forecast-dominant alluvial-plain hazard. Pacific: climatological wetness × current convection. Amazonia: similar pattern at lower population intensity. |
| Indicator | Continuous posterior > 0.5 across the six named regions. |
| Implication | UNGRD's national coordination needs to manage six concurrent regional fronts. Departmental risk-management offices in the named departments should activate municipal coordination. This is not a punctual emergency; it is a country-wide elevated alert. |
| Limits | Single 7-day GFS init; a multi-init or WeatherNext 2 ensemble run would test corridor stability. |
FDO 02 — 12.96 million people in the posterior exposure surface
| Source / trace | CIESIN/GPWv411/GPW_Population_Count.first() at populationScale = 1000 m, posterior-weighted via the expectedPopulation reducer. |
| Simple | About 12.96 million of the model's mapped 32.88 million Colombians live inside the 7-day flood-posterior surface — nearly four in ten. |
| Expert | 12,957,478 expected / 32,884,748 total = 39.40 %. GPWv4.11 vintage is 2010–2015; current Colombia population is ~52 M, so the absolute count is a conservative lower bound. The percentage is the durable finding. |
| Indicator | Population — expected exposed: 39.40 % · 12,957,478 ppl of 32,884,748 ppl total. |
| Implication | This 7-day window should be communicated as a national-scale elevated alert. Departmental civil-protection offices in the eleven densest departments (Antioquia, Atlántico, Bolívar, Magdalena, La Guajira, Arauca, Casanare, Meta, Cundinamarca, Boyacá, Valle del Cauca) should anticipate concurrent activation. |
| Limits | GPWv4 vintage. Percentage robust to vintage; absolute count is a lower bound. |
FDO 03 — Half of the mapped electric grid inside the posterior surface
| Source / trace | Merged predictive grid (sat-io) + Facebook GMV grid, line-painted at 1-pixel width on the exposure grid (lineFracExpo helper). |
| Simple | About half of Colombia's mapped electric grid is inside the 7-day flood surface. |
| Expert | 49.80 % · 27,434 km exposed of 55,090 km mapped. Concentrated along the Andean spine (Medellín–Bogotá–Cali interconnection backbone) and the Atlántico–Bolívar coastal corridor. |
| Implication | XM (national grid operator) and the regional distribution utilities (EPM, Codensa-Enel, EPSA, Electricaribe successors) should anticipate elevated outage risk over the window. Substations in the Sabana de Bogotá, Aburrá Valley, and Atlántico-Bolívar deserve specific pre-positioning. |
| Limits | The Facebook GMV grid is a remote-sensing-inferred network: false positives urban, false negatives rural. The 49.80 % is exposure of mapped lines, not the operational national grid. |
FDO 04 — Three out of every five mapped oil-palm hectares inside the posterior surface
| Source / trace | BIOPAMA/GlobalOilPalm/v1 mosaic; classes 1 and 2 retained; fraction-preserving aggregation at exposureScale = 1000 m. |
| Simple | Roughly three of every five mapped hectares of Colombian oil palm sit inside the 7-day flood surface. |
| Expert | 60.94 % — the highest exposure ratio in the panel and geographically coherent: Colombian palm cultivation concentrates in the Magdalena Medio (Santander–Cesar–Bolívar corridor) and the Llanos plantation belt (Meta, Casanare), both of which sit in the posterior surface. |
| Implication | Fedepalma and large producers (Indupalma, Aceites Manuelita, Palmaceite, others) should anticipate field-access disruption, harvest delays, and oil-extraction-plant inflow disruption over the window. Plantation drainage-canal maintenance crews should pre-deploy. |
| Limits | BIOPAMA is global but date-stamped; recent expansion or contraction of Colombian palm footprint not reflected. |
FDO 05 — Half of mapped urban area inside the posterior surface
| Source / trace | MCD12Q1 LC class 13 (urban), fraction-preserving aggregation at exposureScale = 1000 m. |
| Simple | About half of Colombia's mapped urban area is inside the 7-day flood surface. |
| Expert | 50.20 %. The five largest cities (Bogotá, Medellín, Cali, Barranquilla, Cartagena) each fall partly or wholly inside the surface, alongside intermediate cities (Bucaramanga, Cúcuta, Pereira, Manizales, Armenia, Ibagué, Neiva). |
| Implication | Urban civil-protection in the named cities should activate flood-response protocols. Municipal hospital networks should prepare for arrival of injured and displaced populations from peri-urban informal settlements, where most urban flood mortality occurs. |
| Limits | MODIS urban class is coarse (500 m); micro-flood-prone neighbourhoods below the resolution. |
FDO 06 — Pacific corridor: persistent wetness reads consistently as elevated risk
| Source / trace | Visible posterior over Chocó (Quibdó, Istmina, Tadó) and coastal Valle (Buenaventura) and Cauca. |
| Simple | The Pacific — the wettest region of South America — reads as elevated risk for the full 7-day window. |
| Implication | The Pacific corridor needs sustained, not punctual, attention. Quibdó's chronic urban flooding, the Atrato basin's recurrent inundation cycles, and Buenaventura's port-area flooding remain inside the surface across the window. |
| Limits | Persistent-wetness regions visible in any flood-risk model; the value-added information is the per-basin deviation, not the absolute signal. |
FDO 07 — Amazonia corridor: visible signal at low population density
| Source / trace | Posterior signal across Putumayo, Caquetá, Amazonas. |
| Implication | Departmental risk-management offices in Putumayo, Caquetá, and Amazonas should not be skipped in the regional briefings. Indigenous-territory leaders and the Departamento del Putumayo's risk office deserve direct contact. The 2017 Mocoa avalanche history reinforces the prior over the Andean-Amazon foothills. |
| Limits | 7-day GFS forecast has weaker skill over Amazon convection than over Andean orography. |
Media-triangulation FDOs (UNGRD / IDEAM / regional press)
FDO 08 — Llanos triangulation: model corridor matches active named events
| Source / trace | Model's Llanos corridor (FDO 01) vs.: Arauquita's Public Calamity Decree 100.03-037 (3 June 2026); Cabuyaro (Meta) Río Meta overflow affecting ~700 people; Puerto López Río Metica overflow affecting ~1,050 families; Río Pauto overflow affecting >100 families in Casanare. |
| Simple | The model's Llanos corridor is happening: four named, decreed, or reported flood events already inside it. |
| Expert | Four out of four of the named river-overflow events in the Llanos in early June 2026 fall inside the posterior surface. Arauca and Casanare are in the high-posterior band; Meta sits in the same band along its eastern half. Convergence between model output and ground reality. |
| Implication | Trust the corridor for the remaining window. UNGRD's Llanos liaison and the four departmental risk-management offices should treat the next 7 days as elevated based on combined model-and-ground evidence. |
| Limits | Convergence does not guarantee causal validity — La Niña-conditioned wet-phase models will always over-predict in the Llanos; the same ground events would converge with either baseline. |
FDO 09 — La Guajira triangulation: convergent emergency activations
| Source / trace | Model's La Guajira corridor (FDO 01) vs.: Riohacha alcalde Genaro Redondo's 6 June activation of the permanent Risk Management Committee after a ~2-hour storm flooded multiple barrios; MeteoGuajira flood warning; 8 of 15 La Guajira municipalities (including Maicao) under public-calamity declarations earlier in 2026. |
| Simple | The model's La Guajira corridor matches what local government is already doing: Riohacha's mayor has activated emergency coordination. |
| Expert | La Guajira is the most striking single-departmental convergence between model and ground. Posterior surface covers Riohacha–Maicao–Manaure–Uribia; ground reports describe coordination activations and prior calamity declarations across that same axis. |
| Implication | MeteoGuajira/IDEAM regional office and the Wayuu indigenous-territory leadership in northern Guajira should be the primary information channels for the next 7 days. UNGRD's La Guajira liaison should coordinate with the departmental governor's office. |
| Limits | La Guajira is historically dense in the flood prior; convergence here is partly inevitable. Analyst-positionality confirmation bias risk noted. |
FDO 10 — Andean triangulation: model corridor matches IDEAM red-alert geography
| Source / trace | Model's Andean cordillera corridor (FDO 01) vs. IDEAM 9 June 2026 bulletin (599 municipalities on landslide alert, 141 in red alert, concentrated in Antioquia, Santander, and Chocó); UNGRD's named most-affected departments (Huila, Valle, Antioquia, Cundinamarca, Caldas, Córdoba); Semana / Puerto Triunfo flooding affecting 192 families. |
| Simple | The model's Andean corridor matches IDEAM's red-alert map: Antioquia, Caldas, Cundinamarca, Valle, Cauca all flagged by both. |
| Expert | IDEAM-named red-alert departments (Antioquia, Santander, Chocó) and UNGRD's named most-affected departments all sit inside the model's high-posterior surface. Eight of nine concrete ground-report locations triangulate with the model. |
| Implication | This is the JANUS-method "model-and-ground concordance" finding. Capybara's model independently identified the same priority departments that IDEAM and UNGRD are already monitoring, using independent data and an independent fusion method. That convergence is publishable. |
| Limits | Independence is partial: IDEAM precipitation data feed CHIRPS calibration windows elsewhere in the pipeline. |
Adversarial finding — methodological
FDO 11 — ADVERSARIAL: La Niña baseline mismatched to current ENSO state, yet model still triangulates with ground
| Source / trace | Model run used ENSO La Niña baseline (CHIRPS years 2020, 2021, 2022, 2025). NOAA Climate Prediction Center ENSO discussion June 2026: ENSO-neutral conditions present, with 62 % probability that El Niño emerges JJA 2026 and persists through end-2026. |
| Simple | The model was run with a La Niña reference period, but the current climate state is no longer La Niña. Despite this, the model still matches the ground events that are happening. |
| Expert | The most methodologically informative finding of the run. Three possibilities: (a) mismatched baseline produces a false-positive corridor and ground-convergence is coincidence — weak: six specific corridors, not uniform wet-season pattern. (b) model detects a true wetter-than-La-Niña anomaly that also happens to be wetter-than-neutral — plausible. (c) model is robust to baseline mismatch because the priors (flood archive, embedding similarity) carry most of the spatial structure, baseline mainly tempers magnitude — most likely. |
| Indicator | NOAA CPC: La Niña inactive, neutral with El Niño emergence probable. Model: La Niña baseline. Ground: convergent. |
| Implication | (1) Re-run with ensoPhase = all_years and ensoPhase = neutral to test corridor stability. (2) Operationally: the spatial corridor identification is trustworthy for this window; absolute exposure percentages may be slightly inflated. (3) Framework: ENSO_YEAR_TABLE should be updated to reflect the 2026 transition. Maintenance task, not a model defect. |
| Limits | One run. Repeated runs with corrected baselines across different forecast windows would constitute proper sensitivity analysis. |
FDO 12 — Area-class absolute totals are not quotable in this run; ratios remain interpretable
| Source / trace | Cropland total 742 k ha (FAO documents ~17 M ha for Colombia). Built-up "total" 113 M ha — essentially the entire 114 M ha country area, implausible. |
| Simple | The absolute hectare numbers in the area columns are not internally consistent. The percentages are still valid as ratios. |
| Implication | Quote percentages, not absolutes in public communications until the chain is patched. Apply this rule to FDOs 03–07 in particular. |
| Limits | Defect is in the denominator-aggregation step, not in the posterior risk surface or population reducer. |
FDO 13 — Coverage gaps: model flags Vichada and parts of the Caribbean coast where no fresh June 2026 ground reports surfaced
| Source / trace | Posterior surface includes Vichada, parts of Sucre, parts of Bolívar coastal. Triangulation search for these areas returned no June 2026 named events. |
| Expert | Two readings, both worth holding. (a) Observation bias: Vichada and rural Sucre/Bolívar are exactly the kind of low-population low-media-coverage areas that produce systematic under-reporting. The prior tempering work in the model addresses this asymmetry. Absence of news does not mean absence of risk. (b) Forecast over-call: the corridor could be a forecast artefact. Presence in Vichada specifically — adjacent to the well-corroborated Arauca/Casanare/Meta corridor — slightly favours reading (a). |
| Implication | Volunteers with Vichada or Caribbean-rural ground contacts should specifically check on those zones, treating model flagging as a hypothesis to verify. Cruz Roja Colombiana's regional cells in Vichada and Sucre / Bolívar coastal municipalities are the right channels. |
| Limits | Search may have missed Spanish-language local sources; follow-up Vichada-specific check recommended. |
5. Limitations & caveats
- The ENSO baseline used in this run (La Niña) does not match the current ENSO-neutral state. See FDO 11. Spatial corridor structure remains trustworthy; absolute exposure percentages may be slightly inflated.
- Area-class absolute totals are not quotable in this run (FDO 12). Percentages remain interpretable.
- GPWv4 vintage (2010–2015). Absolute population counts are a lower bound.
- GFS native resolution is 0.25° (~25 km). Sub-municipal precipitation variation not resolved.
- No marine flood component. Coastal sea-wave inundation on either coast invisible.
- Bayesian priors are historical (flood archive ends 2016; satellite-embedding mosaic is 2022).
- N/A rows for oil & gas / wind / solar farms reflect sparse-point sampling artefact, not "zero exposure" findings.
6. AI-use disclosure & reflexive memo (JANUS)
LLM assistance. A large language model assisted with structuring sections, drafting the plain-language Simple text of each FDO, executing the media-triangulation web search across UNGRD, IDEAM, El Tiempo, Semana, La Voz del Cinaruco, and other regional sources, and weaving the convergent findings into the FDO structure. The LLM did not generate the model output, fabricate exposure numbers, or invent the ground-report events; all media-FDO claims trace to the URLs in §7.
Provenance. Exposure figures read directly from the model's bottom-left exposure panel. Model code versioned at PROD/src/earth-engine/integrated-multihazard-model-tuned.js (the Fable-tuned multi-hazard, Bayesian flood class). Median media-source search recency: 4 days. ENSO classification source: NOAA Climate Prediction Center, June 2026 discussion.
Reflexive memo. The analyst is Colombian and a long-time Bogotá resident with operational familiarity with IDEAM, UNGRD, and the named departmental governance structures. This is not an outsider report; care has been taken to flag where insider intuition might be reading concordance into ambiguous evidence (notably in FDO 09's La Guajira finding, where confirmation bias is most plausible).
Adversarial pass. Five alternative explanations considered. (1) Six-corridor structure is a forecast-window artefact — partially accepted as caveat. (2) Ground-convergence is coincidence driven by universally wet first-rainy-season climatology — rejected (six specific corridors, not uniform wetness). (3) La Niña baseline mismatch invalidates the run — rejected (priors carry spatial structure; baseline tempers magnitude). (4) Area-totals defect invalidates spatial findings — rejected (defect isolated to denominator aggregation). (5) Confirmation bias colouring La Guajira finding — accepted as risk; FDO 09 lower-cased in operational priority for that reason.
7. Sources
Model run. Capybara EWS multi-hazard model (Bayesian flood class), PROD/src/earth-engine/integrated-multihazard-model-tuned.js. JANUS analytic framework: Feged 2025, DOI 10.21203/rs.3.rs-7580516/v1.
Ground-report triangulation (June 2026):
- UNGRD (Carrillo): >600 emergencies, 24 departments, 155 k families — UNGRD on X.
- IDEAM 9 June 2026 hydromet bulletin — IDEAM bulletin PDF.
- Arauquita Public Calamity Decree 100.03-037 — La Voz del Cinaruco.
- Cabuyaro & Puerto López Río Metica overflow — El Tiempo.
- Riohacha emergency activation — La Guajira Hoy.
- Puerto Triunfo (Antioquia) — Semana.
- First tropical wave 2026 — Infobae.
- IDEAM 10–12 June outlook — Noticias Caracol.
- IDEAM 2–5 June outlook — El Tiempo.
ENSO state: NOAA Climate Prediction Center, June 2026 ENSO discussion — CPC ENSO Discussion.
Document generated: June 15, 2026 | Model: Integrated Multi-Hazard Model (Fable-tuned) — Flood (Bayesian) | Forecast: GFS short-term ≤168h | ENSO baseline: La Niña (mismatched to current state — see FDO 11) | Bayesian fusion: active | Country profile: standard | Classification: UNCLASSIFIED / FOR PLANNING USE
Labels: Early Warning System, Situation Report, Colombia, Floods, Bayesian fusion, JANUS, UNGRD, IDEAM, La Guajira, Llanos, Caribbean, Andean, Pacific, Amazonia, ENSO transition
Note: Only a member of this blog may post a comment.