capybara geek center

Early Warning Systems - Methods & models

A family of open-source climate early-warning models in constant development.

Capybara is not a single algorithm. It's a family of related models — wildfire, flood, heatwave, and more in development — built on a shared spine of open data, Bayesian fusion, and a publication framework (JANUS) that requires every claim to trace to a source. The family grows when volunteers, partners, or our own analyses surface new gaps to fill.

The shared spine

Every Capybara model is built on the same foundation. When we add a new hazard or refine an existing one, the change inherits the spine automatically — we never duplicate plumbing.

Open data GFS, WeatherNext 2, CFSv2, CMIP6 forecast ensembles. CHIRPS precipitation, SMAP soil moisture, MODIS / VIIRS / FIRMS land surface. Google Satellite Embedding, HydroATLAS, HydroSHEDS rivers. GHSL urban centres, GPWv4 population. GRIP4 roads. All open, all running on Google Earth Engine.	Bayesian fusion (where used) A clean separation between forecast hazard (what the weather is doing) and prior susceptibility (where the place has burned or flooded before, and what it looks like from satellite). The two are fused, calibrated against historical events, and reported with explicit confidence ranges.
JANUS publication framework Every situation report uses the JANUS protocol (Feged 2025) — each claim links to a dataset or source, every report gets an adversarial pass, external triangulation (media, official bulletins) is treated as evidence, AI assistance is disclosed.	Open code on Google Earth Engine MIT-licensed JavaScript on GEE. No installation, free for non-commercial use, runs in a browser. Volunteers and partners can fork, run, and modify the models without any infrastructure setup.

The model family

Status badges: PROD in active use for situation reports · MVP usable but under refinement · PLANNED on the roadmap.

🔥 Wildfire risk model

PROD

Integrated multi-mode wildfire risk surface. Three model classes selectable from the UI dropdown:

Legacy — simple	PML GPP anomaly × SMAP wetness × FIRMS T21 historical fires. Multiplicative collapse → granular per-pixel hotspots. Fastest. Best for public-policy prioritization.
Bayes — fast	Full Bayesian fusion with priors (fire density, satellite embedding) and calibration. Smooth, statistically tidy, fast. Recommended for country-wide overviews.
Ensemble (Bayes + Legacy)	Bayes computes where risk could happen, legacy refines which specific pixels. Slow but granular. Best for detailed reports.

Includes: dynamic fuel-dryness modifier, burnable-fuel mask with one-directional rescue, VIIRS NOAA-20 fire history, four forecast horizons (GFS, WeatherNext 2, CFSv2, CMIP6), and full exposure analysis against population, infrastructure, and protected lands.

💧 Flood risk model

PROD

Integrated multi-mode flood risk surface with two model classes:

Simple — hazard index	Six-component hazard composition (precipitation × slope × nightlights × GSW change × forest loss × soil moisture). Fast, transparent.
Bayesian — full fusion	Hazard fused with flood-density and satellite-embedding priors, calibrated against historical events, with explicit prior tempering against history-dominance.

Includes: per-basin precipitation deviation against ENSO-matched baselines, four forecast horizons, exposure analysis against population, cropland, urban areas, oil palm, protected areas, electric grid, roads, oil & gas, wind, and solar.

☀️ Heatwave risk model

MVP

Heat Index plus baseline-anomaly classification, with explicit shelter identification (buildings, schools, hospitals, cultural centres mapped at high resolution for evacuation planning).

Status: usable for situation reports but under refinement. Coefficients calibrated for tropical-Andean conditions; Mediterranean and continental regimes need empirical re-tuning. Will graduate to PROD once the multi-region validation pass is complete.

🌧️ Drought · 🌊 Hurricane / tropical-cyclone exposure · ⛰️ Landslide

PLANNED

Three hazards on the near-term roadmap. Each one inherits the shared spine — open data on GEE, Bayesian fusion where the science supports it, JANUS publication. Volunteer interest accelerates the timeline; if you want one of these to ship sooner for a region you care about, get in touch.

📄 JANUS — analytic & publication framework

PROD

Not a model — the publication and reasoning layer that sits on top of every model. JANUS specifies how each claim links to a dataset or source, how an adversarial pass is conducted, how external triangulation is integrated, and how AI assistance is disclosed.

Peer-reviewed publication: Feged 2025, DOI 10.21203/rs.3.rs-7580516/v1. Every Capybara situation report cites it.

How a typical run works

1	Open the script. Paste the Earth Engine JavaScript into code.earthengine.google.com. No installation; everything runs in the browser.
2	Pick your scenario. Choose hazard (wildfire, flood, heatwave), country and optional sub-region, forecast horizon (GFS / WeatherNext 2 / CFSv2 / CMIP6), ENSO baseline, and model class. The UI auto-detects country size and picks a sensible spatial profile.
3	Click Run. The output appears as a terrain-shaded posterior risk surface with roads, rivers, and city labels overlaid. An exposure panel populates with the headline numbers — population, infrastructure, protected areas, energy assets at risk.
4	Write the JANUS report. Take a screenshot of the map and the exposure panel, then drop them into the situation-report template. Triangulate against local press and official bulletins. Volunteer reviewers help.

In active development

Visible changes from the past month — the family is in constant flux, and we make every change public.

Recent improvement	What changed
Vertical-exaggeration terrain shading	Posterior surface now renders with SRTM hillshade modulated by 5× vertical exaggeration. Cordillera structure reads through the risk colors at country zoom.
Reference features overlay	Major roads (GRIP4), major rivers (HydroSHEDS), and city centres (GHSL Urban Centres) layered above every posterior map.
Three-mode wildfire dropdown	Legacy / Bayes-fast / Ensemble. Each mode has a clear trade-off between speed, granularity, and statistical tidiness.
Auto country-size profile	The script measures the AOI before running and picks standard, big_country, or huge_country automatically. Mexico, Brazil, USA, Russia, Canada, Australia all runnable.
Sub-region selector	For big countries: pick a state or department (GAUL admin-1) and run at full standard resolution. Brazil → Amazonas, USA → Texas, Mexico → Nuevo León all work.
Single-pass exposure panel	Eleven exposure rows now compute in two evaluate calls instead of twenty-two. National-scale runs return in minutes instead of timing out.

Lineage

The current model family did not appear all at once. It is the synthesis of ~60 prior Earth Engine scripts written between 2018 and 2024 — early FIRMS / MODIS fire readouts, the first SMAP soil-moisture experiments, the original flood-damage panel design, the ENSO-conditioned fire prior, the topological data analysis for epidemics. Each prototype contributed an idea that survives in the current family.

The lineage is documented in full in legacy/LINEAGE.md on the GitHub repository — every legacy script is preserved, labelled with its status (promoted, experimental, abandoned, or scratch), and traced to where its ideas live in the current PROD models. We think this is the right way for an open project to handle institutional memory: every situation report's analytic chain can be followed back to the prior work that informed it.

Use the models yourself

Three steps:

1. Read the CONTRIBUTING guide on the GitHub repo (what we ask, how we behave, how to publish a JANUS-compliant report).
2. Open code.earthengine.google.com and paste the multi-hazard script.
3. Run the model for a country you know. Three things that surprise you become the seed of your first situation report.

Volunteer reviewers walk new contributors through the first run. The technical bar is low — paste a script, pick a country, click Run — and the editorial side is where most volunteer time goes.

What the models can't do

Every Capybara situation report carries a Limitations & caveats section. The family-level limits are worth stating here so they are visible upfront:

• These are proof-of-concept models for situational awareness and planning. They are not regulatory tools. They do not replace official meteorological warnings or operational forecasting systems.
• Forecast skill is the upper bound on accuracy. Even WeatherNext 2's 15-day ML ensemble has irreducible uncertainty. We propagate ensemble spread where the science supports it; we are explicit about what we don't propagate.
• Spatial resolution is limited by the input data. A 25 km forecast field cannot resolve sub-municipal variation; a 90 m DEM cannot resolve sub-90 m landscape detail. We are honest about this in every report.
• Historical priors carry observation bias. VIIRS misses understory fire; flood archives undercount rural events. Our priors include tempering and rescue mechanisms to address this; the residual bias is named in every report.
• Hazard coefficients are calibrated for tropical-Andean conditions, since that is the analyst's geography. Performance for other regimes (Mediterranean, continental, monsoonal, desert) is empirically validated by media triangulation in each report but not formally re-tuned. Volunteer-led recalibration for new regions is an active workstream.

Roadmap

Next 30d	Scheduled daily runs for Colombia, Mexico, Western Europe. Drought model MVP. Heatwave model graduates to PROD.
Next 90d	Hurricane / tropical-cyclone exposure model. Landslide-susceptibility MVP. Formal multi-region calibration of the hazard coefficient set.
Within 2026	Partner programme launch. Volunteer-cohort onboarding beyond Latin America and Europe. Earth Engine App with no-code interface for non-technical users.

---

Capybara Center methods · MIT-licensed open code on GitHub · JANUS analytic framework: DOI 10.21203/rs.3.rs-7580516/v1 · Built on Google Earth Engine