Home/Simulator/Scientific details

The science behind the simulator

The simulator is powered by PyroWISE — a pure-Python, GIS-native, AI-augmented wildland fire-growth engine, built as a clean-room reimplementation of the public, peer-reviewed Canadian CFFDRS / WISE science stack, with every equation kept explicit and citable.

The engine

An operational service, not a one-off model

PyroWISE is not a one-off research model: it is the engine that runs in production behind the Karst Firewall operator dashboard. It pushes a fire front across the real cross-border landscape — fuel, weather, terrain and infrastructure — and, by running many slightly varied simulations, reports the outcome as probabilities (ensemble envelopes and a burn-probability surface) rather than a single predicted line. Every run is reproducible and fully traceable, backed by a scientific kernel under more than 4,000 automated tests.

01 / NOWCAST

If a fire started now

A short-horizon run from live cross-border weather and a single ignition point — where the fire would reach in the coming hours, streamed perimeter by perimeter as it solves.

Live weather · ≤ 48 h
02 / SCENARIO

What-if and planning

Operator-chosen conditions for prevention, training and impact studies — test ignition points, weather and firebreaks, run as a probabilistic ensemble.

Ensemble · up to 10 days
03 / REPLAY

Historical replay

Re-run a recorded fire on the weather it actually burned in — the basis for benchmarking, calibration and after-action review.

Recorded weather · calibration
The scientific core

Peer-reviewed foundations, reimplemented

The physics kernel preserves the established CFFDRS science unchanged — three peer-reviewed systems work together to turn weather, fuel and terrain into a moving fire front.

01 / FWI

Fire Weather Index System

Translates temperature, humidity, wind and rain into fuel-moisture codes (FFMC, DMC, DC) and fire-behaviour indices (ISI, BUI, FWI).

Van Wagner 1987 · NRCan FWI2025 update
02 / FBP

Fire Behaviour Prediction System

Computes rate of spread, intensity and fuel consumption across the 16 canonical fuel types, extended here with a 9-class Karst envelope.

FCFDG 1992 · 16 fuel types
03 / GROWTH

Huygens vector propagation

Advances a polygon fire perimeter through a heterogeneous fuel × wind × slope × barrier field, with self-intersection cleanup and spotting merges.

Tymstra et al. 2010 · NOR-X-417
Calibrated for the Karst

Generic fuel models are not enough

The standard Canadian fuel models are adapted to the Dinaric Karst (the limestone belt running south-east from north-east Italy and Slovenia through Croatia, Bosnia and Herzegovina, Montenegro and Albania — the cross-border Karst is its north-western tip) and NE-Italy vegetation through a 9-class Karst envelope (K01–K09) and a five-grid resolver that fuses species, structure, age, moisture and continuity into the active fuel model per pixel.

  • Versioned priors resolved per scope and season, so production routing is auditable and reproducible.
  • K04 envelope calibration against real events with a CMA-ES / Nelder-Mead optimiser and IoU, Hausdorff and area-error metrics.
  • Ground truth from Copernicus EMS perimeters, Sentinel-2 burn severity and LSA-SAF fire-radiative-power timelines.
  • AI augmentation (a U-Net spread emulator and an ensemble surrogate) layered on top — with the physics kernel always available as fallback and ground truth.
Live fuel state

Reading the landscape's dryness from space

Beyond the static fuel map, PyroWISE is building a dynamic picture of how flammable the landscape is — refreshed from satellites every few hours instead of frozen in a year-old map. The backbone is a familiar number, NDVI (the Sentinel-2 vegetation "greenness" index); the novelty is what the system does with it — comparing today's greenness against what is normal for the season, turning that gap into a vegetation-stress signal, and wiring it straight into the fire-spread physics.

In one line: PyroWISE is turning Sentinel-2 greenness into a live measure of how dry the landscape really is right now — and connecting that, for the first time, straight into how fast it predicts a fire will run. The architecture splits cleanly: PyroWISE owns the science (baselines, anomalies, fuel coupling); the operational web app mirrors and renders it.

From ignition to spread

Two models, one decision

The KFWI service answers where a fire is likely to start and how alarmed to be; PyroWISE answers given an ignition, where the fire goes and by when. They are independent services.

0.934

AUC of the Random Forest ignition model (hotspot density, infrastructure proximity, vegetation, seasonal + diurnal cycle, proper pseudo-absence sampling).

77.5%

fire detection on the 1,227-fire validation cohort for the integrated risk = P(ignition) × severity_weight(FWI) — versus ~29% for FWI alone.

~97%

of Karst fires occur below the generic European EFFIS "High" threshold, so EU defaults under-rate the regional risk — hence the Karst-calibrated thresholds.

KFWI ignition model repository

What goes in

Inputs

  • Live cross-border weather (ARPA FVG OSMER + ARSO), or recorded weather for historical replays.
  • Bilateral terrain (DTM) and canopy height (CHM) from regional LiDAR.
  • Infrastructure as tiered barriers — roads, railways, firebreaks, dry-stone walls (FVG CTRN + SLO RUBIN + OSM) and hydrography.
  • An ignition point, horizon, time step and fuel mapping; plus barrier policy and conservation overlays.
What comes out

Outputs

  • Streaming and final fire perimeters over 2D and 3D terrain.
  • Ensemble envelopes (p10 / p50 / p90) and a burn-probability surface.
  • Arrival-time layers for time-to-impact on points, lines and polygons.
  • Run bundles with full provenance — weather source, AOI profile, barrier policy, fuel classes, calibration confidence — and QGIS-ready exports.
Validation & provenance

Benchmarked against the reference stacks

PyroWISE is a clean-room reimplementation of public science — it benchmarks at the file boundary against WISE / Prometheus (Canadian Forest Service), Cell2Fire and the FARSITE / FlamMap family, computing IoU, Hausdorff and area-error against recorded fire perimeters.

Data & reference feeds

CFFDRS (NRCan)WISE / PrometheusCell2FireFARSITE / FlamMap Copernicus EMSSentinel-2Copernicus C3S / ERA5EUMETSAT LSA-SAF NASA FIRMSARPA FVG OSMERARSOOpenStreetMapNatura 2000ZGSNOAA HYSPLIT

PyroWISE is released under AGPL-3.0, matching the upstream WISE licence. The smoke-dispersion product (NOAA HYSPLIT Gaussian-puff) is currently uncalibrated and shown for context only.