ECOBITS
Forecasting how vegetation fires spread, hour by hour, from NASA open data
nodes in the simulation grid
directed edges between nodes
scenarios per forecast
fire-detecting satellites
Fire moves faster than decisions. EcoBits gets ahead of it.
Traditional tools explain a fire after it is out. EcoBits forecasts where it will be in 3, 6 and 12 hours, the way hurricane models track a storm, from real open data:
| Layer | Source | Resolution | Volume |
|---|---|---|---|
| Active fires | NASA FIRMS: MODIS, and VIIRS on S-NPP, NOAA-20, NOAA-21 | 375 m, 1 km | 630k archived detections |
| Terrain | Copernicus DEM GLO-30 | 30 m, vertical error under 4 m | 16 to 877 m elevation |
| Vegetation & dryness | Sentinel-2 L2A: NDVI, NDMI | 10 m, 20 m | 3 scenes, plus all of Dhofar |
| Weather | Hourly ERA5, Open-Meteo forecast | 9 to 28 km | 3,288 hours, none missing |
| Surface heat & soil | NASA GIBS: MODIS LST, SMAP | 1 km, 9 km | Daily |
| Roads & facilities | OpenStreetMap | Vector | 5,351 industrial sites |
Fire spread as a shortest-path problem on a graph.
nodes (130 × 150 grid)
directed edges
neighbors per node
km², cells of 120 × 113 m
8 adjacent + 8 knight moves
Dijkstra arrival timeO(E log V)
From the ignition point, every node takes the earliest time fire can reach it, using a binary heap.
Elliptical spreadANDERSON 1983
R(θ) = R_head·(1−e)/(1−e·cosθ), with length-to-breadth from mid-flame wind speed.
Grass curingCHENEY 1998
CF = 1.12/(1+59.2·e^(−0.124(C−50))) from NDMI.
Fireline intensityBYRAM
I = 18000·w·R, flame length L = 0.0775·I^0.46.
40 parallel worlds in under a tenth of a second.
Each forecast solves the graph 40 times with perturbed wind, fuel and moisture (Monte Carlo), and restarts every hour from the observed fire perimeter. The blue glow on the model is the probability of arrival.
node solutions per forecast
ms for a full forecast
hour forecast horizons
arrival-time band
Sensitivity analysisMORRIS · 60 RUNS
Effect of each factor on area burned after 6 hours (μ*, km²).
Fire Weather IndexCANADIAN FWI
All 6 components: FFMC, DMC, DC, ISI, BUI and FWI. Matches the reference test case exactly (FWI = 10.10).
Probability calibrationBRIER · BSS 0.67
Over 17 issue times on the synthetic twin, backed by 50 unit tests of the equations.
We fuse NASA detections, filter out false fires, and route the crews.
Sensors
MODIS and three VIIRS satellites, with unified confidence.
Duplicates removed
285 detections down to 188 events (1 km, 60 min).
False fires
110 industrial and 65 offshore, out of 193 events.
Open fires
These are the ones that raise an alert.
Filtering uses NASA's static-source reference (4,445 cells from 630k detections), 5,351 OSM facilities, and recurrence on 3 of the last 7 days.
Crew routingTIME-DEPENDENT DIJKSTRA
Avoids any node the fire could reach before the crew passes, using the earliest 20% of scenarios (P20) plus a safety margin. Crews may cross ground burned over an hour ago; civilians never cross burned ground.
Suppression tactics350 · 1700 · 3500 kW/m
Hand crews, then engines, then air support, then indirect attack, by Byram intensity, slope, access and night. Safety zones: 20+ cells burned over 6 hours ago.
Tested on a real fire, and reported straight.
We searched the FIRMS archive from 2018 to 2026 and found 47 candidates in the Dhofar mountains: 35 industrial, 8 uncertain, 4 open fires. The best one is the Qara fire of 14 January 2019.
VIIRS pixels at 14:18
MW radiative power
km² dNBR burn scar from Sentinel-2
IoU gain from model 1 to model 2, still in development
Leakage guardNO FUTURE DATA
Stops the run if any input comes from after the start time: weather, the fuel image, or detections. The burn scar is used for scoring only.
Synthetic twinINTERNAL CHECK
With actual ERA5 weather: IoU 0.77, precision 0.92, recall 0.82, arrival error 1.45 h. This is internal consistency, not accuracy on a real fire.
To be clear: the model 2 fuel fix was designed after studying this fire, so we need more fires to validate it independently. The real fire was a strip 5 times longer than wide; the model gives 1.6 to 2.9.
One platform, from the satellite to the incident commander.
Operations planOPS PLAN
Safe routing for crews and civilians, a tactic for every suppression sector, evacuation routes, a suggested command post, and a printable briefing.
Science LabLAB
Two synced maps: the observed historical fire against the simulation. Plus historical testing, sensitivity, calibration and the source of every layer.
Live alertsEVERY 10 MIN
New FIRMS fires after false-fire filtering, and high fire weather.
3D map and globeMAPLIBRE
Real terrain from top-down and tilted views, with the fire head track.
Fire heat and frontBYRAM
Flame length, front intensity, and heat at the head, flanks and embers.
Heat and dryness mapsLST · NDMI · SMAP
Land surface temperature, and dryness for all of the Dhofar mountains.
Live weather on the mapOPEN-METEO
Wind particles that vary across the map, plus wind and temperature layers.
Data integrityAR · EN
Every synthetic or estimated layer is labeled. Arabic and English interface.
14 algorithms: Dijkstra · Anderson · Cheney · Byram · Monte Carlo · hourly assimilation · time-dependent Dijkstra · FWI · Horn · Morris · Brier · sensor fusion · false-fire filtering · Marching Squares