AI-Driven Wildland Fire Response System
Agentic decision support for WUI preplanning, fireground intelligence, ICS operations, and resource coordination. Pick a simulated scenario, advance the operational clock, and watch eight specialist agents build the incident picture an Incident Commander would actually use.
Built by an engineer who runs calls: the design borrows ICS structure, Colorado Ready-Set-Go evacuation language, and WUI structure-triage thinking. Every agent output is simulated from structured mock data behind a clean runScenarioAnalysis() seam, so the same UI could later sit on real dispatch, weather, infrared-satellite, and fuels feeds — and push critical weather alerts to crew devices and radio.
System architecture
Ingestion · orchestration · human review · field outputs
The demo runs on mock data, but the shape is production-minded. Feeds normalize into a shared incident state, specialist agents analyze in parallel and synthesize through an IC Copilot, a human IC validates before anything leaves the truck, and outputs land in the formats the fireground already uses.
Data ingestion
Normalize the feeds an incident actually produces.
Fire behavior model
One deterministic model every agent keys into.
Agent orchestration
Eight specialists analyze in parallel, then synthesize.
Human review
The IC is the decision-maker. Nothing auto-executes.
Field outputs
Decision support in the formats the fireground uses.
Data source API catalog: the feeds a production build would sit on
Every public and private API behind the ingestion stage, with current documentation and basic shape. Links verified July 2026. In this build all of these are represented by deterministic demo data except the NIFC fire perimeters, which are real.
Weather: observations and forecasts
- NWS APIPublicNOAA / National Weather Service
REST, GeoJSON-LD. GET /points/{lat},{lon} resolves the grid, then /gridpoints/{wfo}/{x},{y}/forecast and /alerts/active?area=CO. Hourly forecast updates; User-Agent header required, no key.
https://www.weather.gov/documentation/services-web-api - NWS Spot Forecast ProgramPublicNOAA / NWS
Incident-specific request (STQ) in, forecaster-written spot forecast out; GeoJSON feed refreshed ~15 min plus an ArcGIS MapServer of active spots.
https://spot.weather.gov/ - Synoptic Data Weather API (MesoWest)Free keySynoptic Data PBC
REST. GET /v2/stations/timeseries?network=2&token=... (network 2 = RAWS) or /stations/latest; JSON/CSV/GeoJSON; sub-hourly RAWS observations. Free public-data tier, paid tiers above.
https://docs.synopticdata.com/services/weather-api - FEMS (NFDRS, replaced WIMS in 2024)PublicUSDA Forest Service
Read-only REST for hourly NFDRS v4 outputs (ERC, BI, dead/live fuel moistures) and station obs; gov login for write paths.
Exact OpenAPI path not verifiable at review time.
https://www.wildfire.gov/application/fems
Satellite thermal detections
- NASA FIRMS Area APIFree keyNASA LANCE
REST. GET /api/area/csv/{MAP_KEY}/VIIRS_NOAA20_NRT/{w,s,e,n}/{days}; VIIRS 375 m, MODIS, and GOES sources; CSV/JSON; NRT latency ~60-180 min; 5,000 transactions per 10 min per key.
This is the API behind the map’s Active fire (VIIRS) data layer, represented with simulated 375 m detections in this build.
https://firms.modaps.eosdis.nasa.gov/api/area/ - NASA FIRMS US/CanadaFree keyNASA LANCE / USDA Forest Service
Same Area API on a separate US/Canada instance fed by direct-readout ground stations: real-time and ultra-real-time VIIRS/MODIS detections — under a minute after overpass for much of the US and Canada — plus Landsat 30 m detections within ~30 min of overpass.
The initial-attack feed: the global NRT window above collapses to minutes here, and Canada is covered at the same grade. This instance, not the global one, is what a production build over this corridor would ingest.
https://firms2.modaps.eosdis.nasa.gov/usfs/api/area/ - GOES-18/19 ABI L2 Fire (FDC)PublicNOAA (Open Data Dissemination on AWS)
Anonymous S3, NetCDF objects (s3://noaa-goes18/ABI-L2-FDCC/...): fire mask, FRP, temperature per ~2 km pixel; CONUS every 5 min, mesoscale 1 min.
https://registry.opendata.aws/noaa-goes/
Copernicus Sentinel imagery
- Copernicus Data Space Ecosystem APIsFree keyESA / European Commission
Free registration, OAuth2 token. STAC catalogue for search, OData for download, openEO for cloud processing, and Sentinel Hub Process/Statistical APIs (POST an evalscript, get rendered raster or stats). Sentinel-2 revisit ~5 days; free quota, commercial tiers.
https://documentation.dataspace.copernicus.eu/APIs.html
Wind and point forecasts (Windy)
- Windy Point Forecast APICommercialWindy.com / Windyty SE
POST /api/point-forecast/v2 with JSON {lat, lon, model (gfs/ecmwf/iconEu), parameters, key}; JSON time series out; model steps 6-12 h. Trial plus paid Professional subscription.
https://api.windy.com/point-forecast/docs - Windy Webcams / Map Forecast APIsCommercialWindy.com
REST GET /webcams/api/v3/webcams?nearby={lat},{lng},{radius} with x-windy-api-key; embeddable interactive map layers via the Map Forecast API.
https://api.windy.com/webcams/docs
Synoptic guidance and model inputs
- SPC Fire Weather OutlooksPublicNOAA / Storm Prediction Center
Day 1-2 fire weather outlooks (Day 3-8 probabilistic) with Elevated / Critical / Extremely Critical areas; shapefile/KML downloads plus an ArcGIS MapServer; issued twice daily.
https://www.spc.noaa.gov/products/fire_wx/ - NWS NDFDPublicNOAA / NWS
2.5 km CONUS gridded forecast elements (wind, gusts, RH, temperature, sky); GRIB2 via tgftp and AWS Open Data; out to 7 days.
https://vlab.noaa.gov/web/mdl/ndfd - HRRR on AWSPublicNOAA (Open Data Dissemination on AWS)
Anonymous S3 (s3://noaa-hrrr-bdp-pds), GRIB2 plus Zarr; 3 km convection-allowing model, hourly cycles, 18 h forecasts (48 h at 00/06/12/18Z).
https://registry.opendata.aws/noaa-hrrr-pds/ - WFAS / NFDRS fire dangerPublicUSDA Forest Service
Daily national fire danger class and dead fuel moisture maps as rendered images and grids; no formal REST API.
FEMS (under Weather above) is the station-level NFDRS source of record.
https://www.wfas.net/ - NFPORS fuels treatmentsPublicDOI / USGS
ArcGIS REST treatment polygons and points with kind, status, and fiscal year; the production analog of the mitigation-plan overlays this demo reuses.
https://usgs.nfpors.gov/
Incidents and fire perimeters
- NIFC Open Data / WFIGSPublicNational Interagency Fire Center
ArcGIS FeatureServer REST: GET .../FeatureServer/0/query?where=1=1&outFields=*&f=geojson. Current perimeters refresh ~5 min; full interagency perimeter history; incident points via the public IRWIN views.
The one real feed in this build (historic burn scar layer).
https://data-nifc.opendata.arcgis.com/ - IRWINRestrictedDOI / USDA interagency
System-to-system incident data exchange between wildfire applications (CAD, WFDSS, INFORM); GeoPlatform account plus approved integration. Public consumption goes through the WFIGS views above.
https://www.wildfire.gov/application/irwin-integrated-reporting-wildfire-information - InciWebPublicUS Forest Service
RSS 2.0 incident and news feeds plus KML; PIO-entered, irregular cadence (hours). No formal API.
https://inciweb.wildfire.gov/feeds
Fuels and terrain
- LANDFIRE Product Service (LFPS)PublicUSGS / USFS
ArcGIS geoprocessing job API: submit AOI + layer list (FBFM40 fuel models, canopy, slope, aspect), poll job status, download GeoTIFF bundle; products versioned, updated annually or biennially.
This build bakes fuels from an unmodified public FBFM40 stand-in (The Wildland Almanac CONUS v2026.1, CC BY 4.0) and re-points to LANDFIRE with one URL change in the fetch script.
https://lfps.usgs.gov/LFProductsServiceUserGuide.pdf - USGS 3DEP / EPQS elevationPublicUSGS
REST GET /v1/json?x={lon}&y={lat}&units=Meters for point queries; 3DEP dynamic ImageServer/WCS for DEM tiles; lidar-based 1 m where available.
The baked terrain grids in this build come from the Open-Elevation public API (SRTM-derived).
https://epqs.nationalmap.gov/v1/docs
WUI parcels and structures
- County open data (e.g. Jefferson County CO)PublicCounty GIS
ArcGIS Hub GeoServices REST, WMS, WFS; parcel polygons with assessor attributes and the Master Address layer this demo already bakes; typically weekly refresh.
https://data-jeffersoncounty.opendata.arcgis.com/ - Building footprints (Microsoft / Overture)PublicMicrosoft; Overture Maps Foundation
Bulk line-delimited GeoJSON by quadkey (Microsoft Global ML Building Footprints) or GeoParquet on S3 queryable with DuckDB (Overture, which conflates Google + Microsoft + OSM); periodic static releases.
https://docs.overturemaps.org/guides/buildings/
CAD, dispatch, and unit alerting (private integrations)
- Tyler New World / CentralSquare (Zuercher) CADRestrictedTyler Technologies; CentralSquare
Per-agency contracts; CAD-to-CAD interfaces and standard exporter push feeds (XML/JSON on unit status change); integration specs under NDA. This is where dispatch notes and unit statuses originate.
https://www.tylertech.com/products/enterprise-erp/api-catalog - Active911 Open APICommercialActive911 Inc.
OAuth 2.0 with per-agency grants; HTTPS POST to access.active911.com/interface/open_api.php; JSON alerts, personnel responses, and map data pushed on dispatch.
https://github.com/active911/open-api
AVL resource positions
- Geotab MyGeotab SDKCommercialGeotab
JSON-RPC Get/GetFeed on Device and DeviceStatusInfo (lat/lon, speed, heading) with incremental feed tokens; near-real-time. CalAmp Telematics Cloud offers equivalent REST feeds.
In practice AVL often arrives CAD-integrated rather than direct from the telematics vendor.
https://developers.geotab.com/myGeotab/introduction/
Public alerting
- FEMA IPAWS-OPENRestrictedFEMA
CAP v1.2 XML exchange for WEA/EAS origination; requires a signed MOA and alerting-authority status. Archived alerts are public JSON via OpenFEMA (ipaws-archived-alerts-v1).
https://www.fema.gov/emergency-managers/practitioners/integrated-public-alert-warning-system/technology-developers/ipaws-open - Commercial aggregators (Xweather Fires, Ambee)CommercialVaisala; Ambee
REST enrichment feeds: active fires with containment and some perimeters (Xweather /fires/{action}), detections plus risk forecasts (Ambee); useful as cross-checks, not systems of record.
https://www.xweather.com/docs/weather-api/endpoints/fires
Proposed normalized ingestion model: one shape per entity, any source
Source adapters are the only code that knows API formats; agents consume eleven normalized entities. Every record carries a common envelope: source, source_record_id, retrieved_at, valid_at (UTC ISO-8601), geometry (WGS84), confidence (0-1), provenance (raw payload ref), and a dedup_key. Computations run in an equal-area CRS; provenance is preserved so agents can weight and cite sources rather than alter them.
Incident
WFIGS incident locations, CAD, InciWeb, aggregatorsincident_id (IRWIN id canonical), name, type, discovered_at, poo_point, acres, containment_pct, status
Dedup on IRWIN id first, else name + distance + time window; CAD events attach as child dispatch notes until matched.
WeatherObservation
Synoptic (RAWS), FEMS, NWS APIstation_id, network, obs_time, temp_c, rh_pct, wind_ms, wind_dir_deg, gust_ms, fuel_moisture_10h
Timestamps aligned to top-of-interval UTC; dedup on station + time + variable.
WeatherForecast
NWS gridpoints, NWS Spot, Windy Point Forecastlocation, model, issued_at, valid_at, met fields, is_spot, forecaster_text
Never mix issued_at generations in one series; spot forecasts carry incident linkage.
ThermalDetection
FIRMS US/Canada (VIIRS/MODIS/Landsat), GOES FDCdetected_at, sensor, point, frp_mw, brightness_k, pixel_size_m, confidence, day_night
Sensor-specific confidence mapped to 0-1; dedup on sensor + scan time + pixel; clustered for agent use with pixel size kept for weighting.
FirePerimeter
WFIGS current + history (real in this build)incident_id, polygon, perimeter_time, acres_gis, method, is_current
Append-only revisions; area recomputed in an equal-area CRS alongside source acres.
FuelsCell / TerrainCell
LANDFIRE LFPS, USGS 3DEPcell_id, fbfm40, canopy_cover_pct, slope_deg, aspect_deg, elevation_m, landfire_version
Static layers version-pinned per incident and resampled to one lattice so spread models see a single grid.
FuelTreatment
NFPORS, state/county CWPP layerstreatment_id, kind, status, polygon, completed_at, source_plan
Status gates model damping: only completed or active work changes predicted behavior.
Structure
County parcels/addresses, Microsoft/Overture footprintsstructure_id, footprint, centroid, parcel_id, address, occupancy_class, defensible_space_score
Footprints conflated to parcels by spatial join; both sources kept in provenance.
ResourceUnit
CAD export, Active911, Geotab/CalAmp AVLunit_id, call_sign, ics_type, position, heading, status, position_time, assignment
AVL positions and CAD status merged on a unit crosswalk; positions older than 10 min flagged stale.
AlertMessage
NWS alerts, IPAWScap_identifier, event_type, area, sent_at, effective..expires, severity/certainty/urgency, channel, text
CAP is already the normal form; raw CAP XML kept in provenance; dedup on identifier + update chain.
EvacuationZone
County GIS zone layers where publishedzone_id, polygon, status (ready/set/go), status_time, population_est, structure_count, authority
Status transitions kept as event history and joined to AlertMessage when a WEA covers the zone.
What this is, and is not
This is a portfolio demonstration of agentic orchestration applied to a domain I know from the other side of the radio. It shows how a multi-agent system can turn raw incident signal into a structured, reviewable decision-support picture — with confidence surfaced and a human in the loop on every output.
The agents are simulated and deterministic. There is no live inference, no real incident data, and no operational guidance here.
This is a simulated portfolio demo. It is not an official emergency management tool and must not be used for live incident decisions.
Fire behavior, wind-shift mechanics, and spread geometry are modeled after the Lower North Fork (Mar 2012), Quarry (Jul 2024), Cameron Peak (Oct 2020), Hayman (Jun 2002), and Pine Gulch (Aug 2020) fires. Roads, neighborhoods, and hazard profiles follow real geography and the 2021 Elk Creek & Inter-Canyon FPD Community Wildfire Protection Plan; the planning-unit boundaries and relative-risk ratings on the map are the published CWPP plan units (The Ember Alliance / Elk Creek Fire Protection District, with North Fork Fire district draft units), and the evacuation zones follow those real boundaries. In the Pleasant Park and Buffalo Creek scenarios, structure and address points come from Jefferson County open data (Jefferson County Business and Innovation Technologies (BIT) Division - GIS Section and Jefferson County Planning and Zoning Addressing Unit); the incidents, threat statuses, and all per-address assessments are simulated and say nothing about any real property. The commercial hazard sites on the map (propane plants and fuel stations on the US 285 corridor) are real businesses at approximate locations from public listings, but every hazard note and threat status attached to them is simulated occupancy-type doctrine and says nothing about any real business’s condition. Wellington Lake structures, and every callsign and event, are fictional. The historic burn scars available on the map (Buffalo Creek 1996, High Meadows 2000, Hayman 2002, Lower North Fork 2012, and others) are real perimeters from the National Interagency Fire Center InterAgency Fire Perimeter History dataset, and the vegetation baseline under the NDVI reference layer follows real Scott and Burgan FBFM40 fuel-model data (Wildland Almanac CONUS, CC BY 4.0); the Sentinel, VIIRS, and Windy data layers themselves are simulated. The evacuation-status language follows the Ready, Set, Go! program of the International Association of Fire Chiefs. No affiliation with, or endorsement by, any agency, district, company, program, or data provider named in this demo is implied.