Multi-Hazard Nature Intelligence from Live Environmental Data
TerraSentinel standardizes live open environmental data — weather, air quality, river discharge, satellite fire detections — and runs its own explainable risk engines for rain, flood, air, heat, drought, storms, and wildfire. Providers supply variables; TerraSentinel computes the risk scores, localized and risk-ranked for climate-vulnerable communities.
- 6 engines
- self-built, explainable risk models
- +48 h
- forecast horizon on live data
- 100% open
- satellite, weather & air inputs
The problem
The data exists. The last mile of environmental intelligence doesn’t.
Open providers already publish the raw variables — rain rates, river discharge, pollutant concentrations, CAPE, satellite hotspots — hourly, worldwide. But raw feeds arrive as thousands of undifferentiated numbers; a swollen river and a routine tide look the same in a JSON column.
A commune leader needs something else entirely: which hazard matters here, today, how serious is it, and why should we believe the score? Localized, prioritized, explainable — or it isn’t actionable.
Bridging measurement and decision today takes GIS and domain specialists that climate-vulnerable regions rarely have on call. TerraSentinel aims at exactly this last-mile gap — as a decision-support layer, not a replacement for official warnings.
“Providers measure the environment. TerraSentinel decides what it means locally.”
The engines
Six self-built risk engines. One normalized data frame.
External providers supply live environmental variables — never final risk scores. TerraSentinel standardizes them into a Normalized Environmental Data Frame and computes its own scores, modes, confidence, explanations, and actions. Every engine ships its formula, inputs, limitations, and a live/demo run console.
HydroCast
Rain & floodShort-duration rain intensity, multi-day accumulation, ensemble-relative river discharge anomaly and soil saturation — one explainable flood-attention score.
AirSense
Air qualityWHO-anchored risk curves over live pollutant concentrations (PM2.5, PM10, NO₂, O₃) plus stagnation context. Provider AQI is shown only as a benchmark.
HeatShield
Heat stressApparent temperature, wet-bulb physiology, humidity interaction and heat duration — with a hard dangerous-heat override.
DroughtWatch
Drought / drynessA 7-day rainfall-deficit dryness proxy with soil moisture, evaporative demand and VPD. Honest wording: proxy, never an official drought call.
StormCast
Storms & windCAPE, lifted index (CIN-dampened), damaging gusts, probability-weighted downpours and pressure falls — ingredients, never an exact storm path.
FireGuard
WildfireThe original ForestGuard stack — SentinelCore, F-GRADE, F-CAST, spread cones — wrapped as the wildfire engine of the platform.
The solution
A multi-hazard intelligence layer on open data
TerraSentinel turns open environmental measurements into a ranked, explainable operating picture across seven hazards — running free-tier friendly, end to end. FireGuard, the original wildfire module, is one engine of the platform.
Live-data standardization
Open-Meteo weather, CAMS air quality, GloFAS river discharge and NASA FIRMS detections normalized into one environmental frame — standard units, UTC time, per-source quality.
Self-built risk engines
HydroCast, AirSense, HeatShield, DroughtWatch, StormCast and FireGuard compute their own 0–100 scores from raw variables. Provider indexes are benchmarks, never outputs.
Cross-hazard reasoning
Heat×pollution, rain×river, drought×heat×wind — the assessment layer surfaces compounding risks a single-hazard view misses.
F-GRADE scoring
Each detected fire episode gets an explainable 0–100 score with drivers and uncertainty.
F-CAST forecast
Fire-weather risk and wind-forward attention zones from Now to +48 h.
Community verification
Confirm, flag false alarms, mark agricultural burning — feedback feeds evaluation.
The wildfire module
FireGuard · SentinelCore
Evidence-Calibrated Wildfire Decision Engine — one engine of TerraSentinel
FireGuard (formerly ForestGuard) is not a NASA FIRMS clone. NASA FIRMS supplies active-fire detections. SentinelCore turns those detections, forecast weather, land/fuel context, source quality, and uncertainty into explainable local decisions — and now feeds the TerraSentinel multi-hazard assessment as its wildfire engine.
Deterministic state machine
Six decision modes with prohibited claims per mode — elevated fire-weather can never be worded as a detected fire, and a single weak detection can never become an "active episode".
Evidence graph
Every run ships a typed node/edge graph linking detections, forecast, land cover, context, source quality and community feedback to the output — with weighted, reasoned relations.
Evidence-calibrated scoring
Scores carry an explicit calibration method, reliability band and confidence interval. An uncalibrated region says so in the report itself — never an implied validation.
Counterfactual explanations
Each decision names what would change it — humidity, rainfall, corroborating detections — computed from the model’s own documented thresholds, not canned text.
SentinelCore is original to this project — a transparent, inspectable decision model, not a black-box AI detector, and never guaranteed fire prediction.
How it works
From raw hotspot to community-ready alert in four steps
Detect
NASA FIRMS active-fire points (VIIRS ~375 m, MODIS ~1 km) are ingested near-real-time and deduplicated across satellites.
Enrich
Each candidate episode is enriched with forecast weather (wind, humidity, temperature), land-cover fuel context, and regional context packs.
Prioritize
F-GRADE scores detected episodes 0–100 with explanations; F-CAST estimates forward fire-weather risk and wind-forward attention zones.
Deliver
Risk-ranked, explainable alerts reach monitored areas in-app, with optional notification adapters (Telegram, Web Push) when configured.
The scoring engines
Two engines, one honest contract: every score explains itself
Both engines publish their drivers and their uncertainty with every output. A number a community cannot interrogate is not intelligence.
F-GRADE
Active episodesScores detected fire episodes 0–100 so responders triage a ranked queue instead of a wall of raw hotspots.
- P(true fire) — multi-satellite corroboration, confidence, land-cover plausibility
- Episode strength — FRP, detection density, persistence across overpasses
- Spread potential — wind, humidity, vapour-pressure deficit, fuel context
- Exposure & impact — proximity to communities and monitored areas
- Uncertainty penalty — sparse sources or stale weather lower the score, visibly
Satellite detections are not ground-confirmed; false positives (industrial heat, agricultural burning) are possible and are down-weighted, never hidden.
F-CAST
Forecast riskEstimates forward fire-weather risk and wind-forward attention zones from forecast model data.
- Risk score & severity — fire-weather risk 0–100 per horizon: Now → +48 h
- Attention zones — wind-forward cones for where attention should shift next
- Main drivers — each score names the weather variables that produced it
- Uncertainty level — grows explicitly with horizon; beyond +24 h is flagged high
F-CAST estimates fire-weather risk and wind-forward attention zones. It does not guarantee ignition or exact fire spread. F-CAST forecasts fire-weather risk and likely attention zones. It does not guarantee that a fire will occur.
FireGuard SentinelCore
SentinelCore v1.0The Evidence-Calibrated Wildfire Decision Engine composes both engines with NASA FIRMS evidence, Open-Meteo forecasts, regional fuel context and source quality — and shows its state machine, evidence graph, calibration, counterfactuals and full model trace for every run. Not a black box, not guaranteed prediction: inspectable decision support.
Data sources
Built entirely on open, attributable data
No proprietary feeds. Every layer on the command map carries its provider, valid time, and caveat — the same information shown on these cards.
NASA FIRMS
CoreActive-fire detections (VIIRS ~375 m, MODIS ~1 km), near-real-time.
Free MAP_KEY · detections are not ground-confirmed
NASA GIBS
CoreLatest-available daily satellite imagery for the command basemap.
No key · daily imagery lags, never a live camera feed
Open-Meteo
CoreForecast weather grid: wind, humidity, temperature, VPD, precipitation, CAPE, wet-bulb, UV, soil moisture.
No key · model data, uncertainty grows with horizon
Open-Meteo Air Quality (CAMS)
CorePollutant concentrations (PM2.5, PM10, NO₂, O₃, SO₂, CO, dust) for AirSense. Provider AQI kept as benchmark only.
No key · grid model, street-level exposure may differ
Open-Meteo Flood (GloFAS)
CoreDaily river discharge + ensemble statistics for HydroCast river anomaly.
No key · ~5 km river grid, ensemble-relative reference
ESA WorldCover
Context10 m land-cover map for fuel classification and plausibility checks.
Static 2021 map · conditions may have changed
RainViewer
OptionalRadar reflectivity frames (past + short nowcast) where coverage exists.
Optional · coverage not guaranteed everywhere
Supabase · PostGIS
PlatformGeospatial data brain: episodes, scores, areas, alerts, ingestion runs.
Row-level security · service keys stay server-side
OpenStreetMap-compatible basemap
PlatformDark cartographic base tiles (CARTO / OSM attribution preserved).
No key · community-maintained cartography
Evidence
Designed to be checked, not believed
Deterministic demo & replay
A bundled scenario dataset drives the full pipeline with zero credentials, and replay controls appear only when a real historical archive has been imported.
Source-quality tracking
Every layer carries provider, valid time, latency, and resolution. Stale or missing sources are labelled — a LIVE badge cannot appear without a recent successful ingestion.
Backtesting-ready architecture
Historical FIRMS and reanalysis weather can be imported and scored against recorded outcomes; evaluation metrics are exposed on the Evidence page.
No fake real-time controls
Timeline controls are generated from data that actually exists. Horizons the model did not serve render disabled, with the reason.
Ethics & safety
Honest boundaries, by construction
Decision support, not command
TerraSentinel is a decision-support tool and does not replace official emergency services, government warnings, or professional environmental management decisions.
Forecasts stay forecasts
F-CAST estimates fire-weather risk and wind-forward attention zones. It does not guarantee ignition or exact fire spread.
No fake real-time
The interface can only claim LIVE when a recent ingestion actually succeeded. Demo, stale, and forecast states are labelled exactly as what they are.
Uncertainty is visible
Every score ships with its main drivers and uncertainty factors. Sparse satellite sources or stale weather lower confidence — visibly, never silently.
Community verification
Local observers can confirm fires, flag false alarms, and mark agricultural burning; feedback is recorded for evaluation and model improvement.
Privacy-light by design
Monitoring areas need no personal profile. Optional notification adapters are strictly opt-in and the platform works fully without them.
TerraSentinel is a decision-support tool and does not replace official emergency services, government warnings, or professional environmental management decisions. Always follow guidance from local authorities and official emergency channels.
See the operating picture
The Multi-Hazard Assessment Lab and Risk Map run in clearly-labelled demo mode with zero credentials — and switch to live open data (Open-Meteo, CAMS, GloFAS, NASA FIRMS) the moment it is reachable.