Simulation · Geospatial · Decision Support
A live-data wildfire simulator that fights real terrain, real buildings, and real weather — then races autonomous drone doctrines against identical fires to find out which response would have saved the most.
Overview
Pick any point on Earth and any historical date range, and Wildfire Sentinel assembles a working simulation of that place: elevation from terrain tiles, building footprints and critical facilities from OpenStreetMap, water bodies, and hour-by-hour reanalysis weather for the dates you chose. No API keys, no backend, no bundled fixtures — every input is fetched live from open sources in the browser.
It started as a hackathon prototype that projected fire spread onto a hand-drawn PNG of Pacific Palisades. The rebuild replaced every part of that: real terrain instead of an image, a physically-grounded spread model instead of weighted coin flips, and a comparison harness instead of a single playthrough.
The capability that makes it a decision tool rather than a visualization is the Strategy Lab. It clones a paused incident, races response doctrines against seeded-identical fires, and ranks them by what they actually saved — so the answer to "what should we do right now" is measured rather than argued.
Architecture
A scenario loader fetches terrain, infrastructure, weather, and the observed perimeter in parallel from four keyless public APIs, with a fallback for each, and rasterizes them onto one grid. The fire engine, drone planner, and Strategy Lab all run in the browser; Monte Carlo ensembles are pushed onto a Web Worker so the map keeps animating while they run.
Pipeline
01
A scenario is a center point, a radius, and a date range — everything downstream is derived from those three inputs.
02
Four public sources are fetched in parallel and rasterized onto a common grid, with fallbacks for every one of them.
03
The engine advances fire by accumulating arrival progress toward each neighbour at a physical rate of spread.
04
Drone stations dispatch under a pluggable doctrine each planning cycle, laying retardant ahead of the front.
05
An ensemble projects the fire forward against the real wind timeline to map burn probability and expected loss.
06
The Strategy Lab clones the paused incident and runs doctrines head to head, then ranks them by value saved.
Fire model
A fire that spreads at the same rate in every direction draws a circle, which is the one shape wind-driven fires never make. Computing rate of spread per direction — fast with the wind, slow across it, almost nothing against it — is what makes the burn footprint resemble the real thing.
Approach
marks a tuned value withheld on purpose. Each technique is described in full; the specific constants that took experimentation to find are not published.
The prototype rolled a weighted die per cell per hour, which meant spread rate was a tuning knob with no physical meaning. The rebuilt engine has every burning cell push progress into its neighbours at a rate of spread in metres per hour, and a neighbour ignites once the front has actually covered the distance to it. Spread speed now denominates in real units, which is the precondition for validating against a real fire at all.
Rate of spread is computed per direction rather than isotropically: the head runs hard downwind at up to ▓▓▓ the base rate, flanks creep, and backing spread is near zero. This is what produces the cigar-shaped burns real wind-driven fires make instead of the expanding circles a naive model draws.
Dead fuels dry and re-wet on a time constant of about ▓▓▓, layered over a multi-day rain memory. Fires therefore stay dangerous into the evening after a hot afternoon and stay suppressed for days after a storm — behaviour that a model reading only the current hour's humidity cannot produce.
Strong winds loft embers downwind on an exponentially distributed distance, and they only take hold where fuels are dry enough to receive them. This is the mechanism by which real fires cross rivers, motorways, and retardant lines — without it, any containment strategy looks far more effective than it is.
Comparing response strategies against separately randomized fires measures noise. Every clone in a Strategy Lab race is seeded identically, so all doctrines fight the same fire, ember for ember, and the difference in outcome is attributable to the strategy rather than to luck. The comparison harness is the product; the fire model is what makes it credible.
The same planner drives the running simulation and the offline races, so a doctrine that wins an experiment is the exact code that then takes over the live incident. Diverging the two would make every experimental result a claim about a system nobody is running.
Terrain and weather land first so the map is interactive in about a second; building footprints stream in behind it across mirrored endpoints and repaint as they arrive. Everything is cached locally, so returning to a region is instant. A tool that takes thirty seconds to become useful does not get used during an incident.
Strategy Lab
Pause a live incident and the lab clones it once per response doctrine. Because every clone is seeded identically, they all fight the same fire down to the ember, so the gap between outcomes measures the strategy rather than the weather it happened to draw.
Rigor
The default scenario replays January 2025 with the recorded ignition point and the actual Santa Ana wind timeline. An uncontained run destroys roughly 7,500 structures against about 6,800 in the real event, and reproduces the observed pattern of wind-driven westward runs punctuated by stalls at high humidity.
That agreement comes with no per-fire tuning — no fitted constants for this incident. Being roughly right without being fitted is what licenses the comparative claims: if the model tracks reality unaided, then "doctrine A saves more than doctrine B" carries real information.
Damage forecasts run a Monte Carlo ensemble whose members diverge through wind variability, producing a burn-probability field and an expected-loss curve rather than one confident trajectory. A single projected outline would imply precision the underlying physics does not have.
Races are ranked on structure cells lost rather than on area, so a doctrine that gives up scrubland to hold a town ranks correctly instead of looking like a failure. Expected loss is also reported value-weighted, with critical facilities carrying weights well above ordinary buildings, and two of the doctrines use that value directly when choosing what to defend.
Trade-offs
Every data source is a key-less public API, so there is nothing to provision, no secret to leak, and no server to keep alive. The cost is working within CORS and rate limits; the benefit is that the whole thing deploys as static files and still runs years from now.
Shipping a canned dataset would have been faster and would have made the tool a demo of one fire. Fetching live means any location and any date since the reanalysis record begins, which is the difference between a case study and an instrument.
Each input has a fallback — a second elevation provider, mirrored infrastructure endpoints, a land-use estimate when building queries fail, bundled weather of last resort. An incident tool that returns nothing when one upstream is down is worse than one that degrades and says so.
The hillshaded base is drawn once per scenario and the dynamic layer paints only the cells whose state differs from it, which for most of a run is the fire front, the scar, and the treated cells. The headroom that buys is spent on ensemble forecasting, which is a better use of the frame budget than redrawing static terrain.
Built with
Reference
Thomson, K. (2026). Wildfire Sentinel: AI Wildfire Simulator on Real Terrain. Technical case study, kylanthomson.dev. https://kylanthomson.dev/projects/wildfire-sentinel (updated Sep 2026).