FLOOD RISK SITUATION REPORT - El Salvador
El Salvador | GFS Short-Term Forecast | El Niño Baseline | June 2026
Integrated Bayesian Flood Risk Model — Google Earth Engine | 7-day Forecast Horizon | Flood-density + Satellite-Embedding Priors + Naïve Calibration + Log-Odds Fusion
| HORIZON | GFS short-term (≤168 h) | ENSO | El Niño (2019, 2023, 2024) |
| FUSION | Bayesian (prior × hazard) | RES | 500 m fine / 5 km basin / 1 km pop |
1. Executive summary
This run replaces the v1 hazard-only model with the Bayesian fusion variant: flood-density prior built from large_flood_events_1985-2016 event points, satellite-embedding cosine-similarity priors from the same flood events and an enlarged all-hazard set (gdis_1960-2018 + landslides), naïve calibration with Laplace-smoothed binning, and log-odds posterior fusion with the 7-day GFS forecast hazard. The ENSO baseline was set to El Niño (years 2019, 2023, 2024).
The spatial pattern is materially different from v1. The dominant flood corridor moves east, anchored on the lower Lempa and Río Grande de San Miguel basins, with a strong central-north arm through the upper Lempa (Chalatenango, Cabañas, Cuscatlán) and a smaller spur in Cuscatlán/Suchitoto. Population exposure is now reported (the v1 reducer artefact is resolved): 22.48 percent of mapped population is in the posterior exposure surface — roughly 1.05 million people of 4.66 million total.
Area-class totals (cropland, urban, protected, built-up) are implausibly small in this run. The numbers are arithmetically consistent (percentages × totals ≈ expected) but the underlying totals are off by ~1000× compared with v1. The cause is the area-aggregation chain (forceDefaultProj + reduceResolution + reproject + reduceRegion) which collapses on small AOIs. This is documented in FDO 09 and §6; the relative percentages within this run remain interpretable, but absolute hectare counts should not be quoted until the chain is patched.
The Bayesian fusion is doing visible work: the eastern concentration is consistent with the historical flood-event clustering in the Lempa system (the priors pull the posterior toward those zones). Compared with v1's purely hazard-driven western-corridor signal, v2 is a structurally different reading of the same forecast.
2. Exposure analysis
Posterior-weighted expected exposure inside the El Salvador AOI for the 7-day GFS window under the El Niño baseline, with Bayesian fusion active. Heatmap cells show the % of total. Greyed-out cells are N/A or implausibly small; see FDO 09 and §6.
| Variable | % of total | Expected (absolute) | Total (denominator) |
|---|---|---|---|
| Population (expected exposed) | 22.48% | 1,046,811 ppl | 4,656,532 ppl ✓ |
| Cropland area | 10.01% | 19.83 ha ⚠ | 198.08 ha ⚠ |
| Urban area | 31.90% | 3.78 ha ⚠ | 11.84 ha ⚠ |
| Oil palm area | N/A — no mapped oil palm in El Salvador | ||
| Protected areas | 18.69% | 39.04 ha ⚠ | 208.86 ha ⚠ |
| Built-up footprint | 22.32% | 258.32 ha ⚠ | 1,157.30 ha ⚠ |
| Electric grid length | 23.80% | 14.99 km ⚠ | 63.00 km ⚠ |
| Roads length | 22.91% | 373.70 km | 1,631.41 km ⚠ |
| Oil & gas sites | N/A — sparse-point artefact | ||
| Wind farms | N/A — sparse-point artefact | ||
| Solar farms | N/A — sparse-point artefact | ||
Robust headline: Population exposure of 22.48% (≈1.05 M people of 4.66 M total) is the most defensible number from this run. The other area-class percentages are stable but the absolute hectare numbers cannot be quoted because the denominators collapsed (see FDO 09).
3. Spatial findings by zone
| Eastern corridor — CRITICAL | San Miguel, Morazán, La Unión, eastern Usulután. Anchored on the Río Grande de San Miguel basin, lower Lempa, and Gulf of Fonseca coastal plain. Dominant signal driven by historical flood-event clustering (the Bayesian prior). |
| Central-north corridor — HIGH | Chalatenango, Cabañas, central Cuscatlán, San Vicente. Upper Lempa basin and the Cerrón Grande / 5 de Noviembre reservoir cascade. |
| Central — MODERATE | La Libertad (east), San Salvador, La Paz (west). Patchy moderate signal on the southern edge of the central-north corridor. |
| Western — LOW IN THIS WINDOW | Ahuachapán, Santa Ana, Sonsonate, western La Libertad. Below detection under El Niño + Bayesian. Not necessarily safe — see FDO 07. |
4. Forecasting and Decision Objects (FDOs) — JANUS method
FDOs follow the JANUS analytic method (Feged 2025, DOI 10.21203/rs.3.rs-7580516/v1). Every assertion below links to a specific dataset, model layer, or visible feature of the output map.
FDO 01 — Bayesian fusion shifts the dominant corridor from west (v1) to east (v2)
| Source / trace | v1 hazard-only output (la_nina baseline); v2 Bayesian fusion output (el_nino baseline) — both visible as the posterior risk layer in their respective runs. |
| Simple | Adding historical flood data to the model moves the predicted danger zone from western to eastern El Salvador. Both readings are defensible — they describe different scenarios. |
| Expert | The Bayesian priors weight the posterior toward historical flood-event density and toward landscapes whose Google Satellite Embedding vectors are similar to those at past flood centroids. El Salvador's historical flood-event clustering is in the Lempa basin and the Río Grande de San Miguel. |
| Indicator | Visible posterior shift: v1 western corridor → v2 eastern + central-north corridor. |
| Implication | Treat the two runs as complementary scenarios. Decision-makers should consider both when allocating resources. |
| Limits | Single-run comparison; not yet replicated across multiple GFS init times. |
FDO 02 — One in five residents is in the posterior exposure surface
| Source / trace | CIESIN/GPWv411/GPW_Population_Count.first() at populationScale = 1000 m, posterior-weighted via the fixed expectedPopulation function. |
| Simple | About 1.05 million of the model's mapped 4.66 million Salvadorans live inside the 7-day flood-posterior surface. Roughly one in five. |
| Expert | 1,046,811 expected / 4,656,532 total = 22.48 percent. GPWv4.11 underestimates current El Salvador population (~6.5 M nationally) because the dataset vintage is 2010–2015. The percentage is the durable finding; the absolute count is a lower bound. |
| Indicator | Population — expected exposed: 22.48% · 1,046,811 ppl. |
| Implication | Civil Protection and affected municipal governments should treat this as a country-wide alert with eastern + central-north emphasis. |
| Limits | GPWv4 vintage (2010–2015). For current absolute counts, swap in WorldPop. |
FDO 03 — Eastern lowlands dominate the exposure: lower Lempa + Río Grande de San Miguel
| Source / trace | Visible map signal (purple/dark blue) in eastern El Salvador; Bayesian prior driven by large_flood_events_1985-2016 whose centroids cluster in the eastern lowlands. |
| Simple | The model concentrates flood risk in the eastern half of the country — San Miguel, Morazán, La Unión, eastern Usulután. |
| Expert | The flood-density component dominates in this region. The Lempa lower basin and the Río Grande de San Miguel have recurrent flood-event reports in the 1985–2016 archive. Embedding similarity reinforces the pattern. |
| Indicator | Visible posterior > 0.5 covering most of San Miguel + Morazán + La Unión + eastern Usulután. |
| Implication | Departmental Emergency Commissions for San Miguel, Morazán, and La Unión should integrate this 7-day window into their pre-positioning. |
| Limits | The prior amplifies where past floods occurred. Areas with no historical record may be under-weighted. |
FDO 04 — Central-north corridor: upper Lempa basin (Chalatenango, Cabañas, Cuscatlán)
| Source / trace | Visible map signal across Chalatenango, Cabañas, Cuscatlán; HydroATLAS basin geometry; CEL reservoir cascade. |
| Simple | A second arm of risk crosses the upper Lempa basin in the central-north — through Chalatenango, Cabañas, and into Cuscatlán. |
| Expert | Couples basin-deviation in the GFS forecast with embedding similarity to the lower-Lempa flood archive. Cascading effects on the CEL hydroelectric reservoir system merit specific attention. |
| Indicator | Continuous posterior signal Chalatenango → Cabañas → Cuscatlán. |
| Implication | CEL should treat this 7-day window as elevated for reservoir-cascade management. Coordination with downstream civil-protection offices. |
| Limits | Model does not represent dam operation. Real reservoir state may modulate the exposure pattern. |
FDO 05 — Population reducer fix: verified
| Source / trace | expectedPopulation() rewritten with explicit reproject({crs, scale}) of both inputs before multiplication, and explicit crs, scale in the reduceRegion call. v1 returned N/A; v2 returns 1,046,811 / 4,656,532. |
| Simple | The bug that made the previous run say "Population: N/A" is fixed. The new run returns a sensible population number for the first time. |
| Indicator | v1: N/A. v2: 22.48% · 1,046,811 ppl / 4,656,532 ppl. |
| Implication | The same fix pattern can be ported to other models where a similar artefact appears. |
FDO 06 — El Niño baseline isolates structural susceptibility
| Source / trace | CHIRPS baseline filtered to El Niño years (2019, 2023, 2024). NOAA CPC ENSO status (June 2026) is El Niño Watch with ONI ≈ +0.9 °C. |
| Simple | The baseline is dry-year (El Niño). Any wet weather in the forecast looks like a big positive deviation against this dry baseline — appropriate for the current ENSO state. |
| Implication | Communicate that v2 is "what does a wet event look like compared with a typical El Niño dry year." A future run with ensoPhase = all_years would identify which corridor remains regardless of baseline. |
FDO 07 — Western corridor (v1) is below detection in v2: not necessarily safe
| Source / trace | v1 posterior surface vs v2 posterior surface for the same departments. |
| Simple | The western corridor that dominated the previous run is not flagged in this one. That doesn't mean those departments are safe — it means under different assumptions, the dominant signal is elsewhere. |
| Implication | For operational decisions, present both scenarios. Tropical Storm Cristina (June 2026) had documented impact in the west; v2 under-weights this because of its preference for historical clusters. Do not interpret v2 silence as a clearance. |
FDO 08 — Bayesian priors are doing visible work — not a re-skin of the hazard field
| Source / trace | The visible eastern + central-north signal in v2 does not match the 7-day GFS precipitation distribution alone. It requires the flood-density prior. |
| Simple | The model isn't just showing where it's raining. It's showing where rain plus historical flood vulnerability combine. |
| Implication | The Bayesian fusion is methodologically active. Reports can cite both the forecast-only signal (p_like) and the posterior (p_post) — they are different objects. |
FDO 09 — Area-class denominators collapsed: known bug, ratio still usable, absolutes not citable
| Source / trace | v2 cropland total: 198 ha. v1 cropland total: 353,300 ha. Difference factor ~1782×. Same for urban (1616×), protected (~1928×), built-up (~2576×). |
| Simple | The "of X total" numbers in the area columns are dramatically smaller than they should be. The percentages remain valid but the absolute hectare numbers should not be quoted until the bug is fixed. |
| Expert | The forceDefaultProj + reduceResolution + reproject + reduceRegion chain inherited from the legacy code collapses for small AOIs (~21,000 km²). The chain works for large countries (Colombia ≈ 1.14 M km² gave sensible numbers). Cause is likely autoAggScale selecting 20 km coarse cells, leaving too few cells covering El Salvador. |
| Indicator | Cropland total 198 ha vs FAO-documented ~700,000 ha cropland in El Salvador. Off by ~3500×. |
| Implication | Quote percentages, not absolutes until the patch lands. The fix is mechanical: revert area-helper functions to the direct-reduceRegion pattern. ~30 minutes including a re-run validation. |
| Limits | Identified inside this run; the model defect is patchable. Bayesian fusion logic, population fix, and spatial corridor identification are all unaffected. |
5. Limitations & caveats
- Area-class absolute totals are not quotable in this run (FDO 09). Percentages remain interpretable.
- N/A rows for oil palm, oil & gas, wind, solar reflect sparse point assets combined with
pointSampleScale = 250 m. Not "zero exposure". - GPWv4 vintage (2010–2015). Absolute population counts are a lower bound.
- GFS native resolution is 0.25° (~25 km). Sub-municipal precipitation variation is not resolved.
- No marine flood component. Coastal damage from sea-wave inundation (Acajutla during Tropical Storm Cristina, June 2026) is invisible to this model.
- ENSO 2025 La Niña and 2026 transition are not yet encoded. A sensitivity run with
all_yearsis recommended before high-stakes use. - Bayesian priors are historical (flood archive ends 2016). New event clusters since 2016 do not contribute.
- The 6-component hazard coefficients were calibrated for Colombia and not re-tuned for El Salvador.
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, and clarifying phrasing. The LLM did not generate the model output, fabricate exposure numbers, or invent dataset names.
Provenance. All exposure figures are read directly from the model's bottom-left exposure panel. The model code is versioned at DEV/src/earth-engine/floods-slides/integrated-flood-risk-model.js (Bayesian fusion edition, 1208 lines).
Reflexive memo. The analyst is an outsider to Salvadoran flood-response operations and has not consulted MARN, Protección Civil, CEL, or municipal civil-protection offices for ground-truth corroboration at the time of writing. The FDO implications should be treated as analytic suggestions, not operational orders.
Adversarial pass. Three alternative explanations were considered: (1) the eastern concentration could be archive bias — partially accepted; the embedding-similarity prior independently reinforces the same region. (2) Bayesian fusion could be over-confident due to calibration on the same archive — accepted; calibration strength is 0.6, not 1.0. (3) The area-totals bug could mean the spatial signal is also broken — rejected; the bug is in the denominator-aggregation chain, independent of the posterior surface.
Document generated: June 11, 2026 | Model: Integrated Bayesian Flood Risk Model v2 (GEE) | Forecast: GFS short-term ≤168h | ENSO: El Niño (2019, 2023, 2024) | Bayesian fusion: active | Country profile: standard | Classification: UNCLASSIFIED / FOR PLANNING USE
Labels: Early Warning System, Situation Report, El Salvador, Floods, Bayesian fusion, El Niño, GFS
Note: Only a member of this blog may post a comment.