WLF · Physics
Wildfire
Cell2Fire fire-spread simulations on a CC-BY global fuel map, emulated to 90 m, anchored to 23 years of NASA MODIS burned area (2001–2023) with a per-cell local history anchor and calibrated so the global level matches the 2014–2023 mean · v1.0 · 2026-09
Western US — Sierra Nevada, Southern California chaparral and the Great Basin at 90 m
Physics-based wildfire likelihood: fire-spread simulations emulated to 90 m and anchored to 23 years of MODIS burned area. Tested against independent fire inventories — MTBS (US) AUC 0.88, NBAC (Canada) 0.92, EFFIS (Europe) 0.82 — with pixel-level AUC 0.95 against 20 years of 500 m MODIS burn history and rank agreement 0.52 with the USFS FSim model. Burn probability is relative to MODIS-detectable burned area, so small fires are under-detected; built-up land scores 0 by construction — use the exposure band for addresses; placement inside a 0.25° cell has modest skill (within-cell AUC 0.61); and it is a long-run annual probability, not a post-fire state — a just-burned area is not lowered. The downloadable rasters carry the ANNUAL probability; the API /score endpoint returns the 50-year probability by default (?horizon=50|100|250|500, or ?band=susceptibility for the annual value) — same product, two bases, so compare like with like. Refreshed annually against the burned-area record, cut each spring once the prior year of burned-area files is complete.
What ships
Surfaces — six 90 m global bands on land 60°S–75°N: annual burn probability per pixel (0–1; built-up, water, bare and ice read 0), exposure (highest annual burn probability within 2 km — what an address in a town beside wildland sees), a global five-class cut and a region-relative one, uncertainty, and coverage (non-burnable / burnable / cropland). Returns — probability of at least one burn within 50 / 100 / 250 / 500 years (rp50–rp500, 0–1).
Every release: 0–1 score · 5-class band · per-pixel uncertainty · input-coverage mask · STAC item · methodology doc.
Inputs
ESA WorldCover 2021 + Petrel fuel crosswalk v1.4 (Scott & Burgan FM40 classes), Petrel Ignition Layer v0.3, ERA5 fire-season weather climatology, Copernicus GLO-30 terrain, NASA MODIS MCD64CMQ burned area 2001–2023 (anchor), Cell2Fire simulation campaign (28,498 fires)
Version history
What changed between releases — the same notes that ship with the data. Newest first.
Surfaces
v1.0
2026-09
First shipped release of the Petrel wildfire product. An earlier statistical line was listed as v1.1 but never shipped (its entry is kept below for the record); several internal builds of this physics line preceded it.
- Simulation-emulation model at 90 m. Ignition-centric Cell2Fire fire-spread simulations over a designed global sample — 2,261 one-degree clusters stratified across 68 fire-region × land-cover strata, 33,082 seed fires — feed a stopping model calibrated to Global Fire Atlas durations and a gradient-boosted footprint emulator (holdout R² 0.80 on log fire area). Expected fire size is evaluated per (0.1° cell × fuel model × slope) stratum on the 90 m fuel grid; fires are placed on burnable 90 m pixels and their footprints convolved analytically at the scale each fire-size class resolves. The 90 m structure follows fuel, terrain and the wildland edge — a built-up pixel reads 0 beside a scored forest ridge.
- Fire rate from the Petrel Ignition Layer v0.3. Fires per year per cell from the anthropogenic ignition probability (inverted to expected occurrences) and the lightning ignition index, with dryness, precipitation, canopy, temperature and elevation covariates. The rate model is fitted on v0.3 itself — the lightning index enters through log1p, so v0.3's rescale of that band needed a refit, not an intercept adjustment.
- Anchored to NASA MODIS MCD64 burned area, 2001–2023. Three steps: a covariate-space correction fitted at 0.5° separately within each of the 14 GFED basis regions (region definitions only — no GFED burned-area data is used) and applied as a smooth 0.1° field; then a local history anchor that sets each 0.25° cell's level from its own 23-year record by credibility weighting (Gamma-Poisson posterior mean, prior weight chosen by a temporal holdout — calibrate 2001–2012, score 2013–2023 — and the log-ratio interpolated bilinearly so the anchor grid is not printed on the product); then the global level set to the 2014–2023 mean, the latest decade (3.64 Mkm²/yr against 4.03 for the full record). Held-out Poisson deviance 0.43 of the covariate-only baseline (0.1° prototype: 0.20).
- CC-BY fuel crosswalk v1.4. WorldCover 2021 × Köppen-Geiger → FM40, with Copernicus LC100 bare-cover fraction deciding burnable vs bare on sparse land (replaces climate-map stair-steps on desert margins), LC100 tree cover and Meta/WRI canopy height v2 for canopy structure. v1.4 splits Mediterranean-climate tree cover by canopy height (below 6 m SH5 chaparral, 6 m and above TU5 timber-understory) and moves tropical rainforest from TL9 to TL2 low-load timber litter. Cropland is burnable (GR2) and flagged as coverage 2.
- Surfaces. Susceptibility is the annual burn probability 1 − exp(−λ); classification is quintiles of λ over burnable pixels, globally and (classification_regional) within the pixel's fire region; uncertainty is a model-internal log-scale σ of λ (1.0 = unconstrained); exposure is the annual probability of a wildland fire within 2 km, defined on non-burnable land; coverage is 0 non-burnable / 1 wildland fuel / 2 cropland. Non-burnable land reads 0, never nodata.
- Validation. MODIS burned area is a calibration input at the 0.25° cell level, so agreement with it at that scale is not quoted as validation. The product is scored out of time on the MODIS record (2013–2023 holdout), within cells against MCD64A1 500 m burn counts, and against records independent of the anchor — MTBS, NBAC, EFFIS burnt areas 2016+, ESA Fire_cci — plus the USFS FSim burn-probability benchmark for the conterminous US. Results in model_card.md.
- Known limits. Spread simulated at 250 m (the campaign ran on the ~270 m v1.1 crosswalk; the 90 m structure comes from the strata, not from 90 m simulation); ERA5 fire-season climatology cannot resolve Santa Ana / Diablo / föhn events; MCD64 is a 500 m product that under-detects small fires, so prescribed, understory and agricultural burning are under-represented and low rates in those regimes are relative to MODIS-detectable burned area; the anchor adds no structure finer than 0.25°; no intensity or flame-length products in this vintage (horizon burn probabilities ship as the companion Returns line). See model_card.md.
- Versioning and annual refresh. v1.0 is a conventional product version — the first release of this model line. The next refresh is v1.1, or v2.0 if the method changes; the label is not a year and carries no data currency, so the currency is stated separately and prominently: burned-area record 2001–2023, global level 2014–2023, WorldCover 2021 fuel through crosswalk v1.4, Ignition Layer v0.3. The product is refreshed once a year, cut each spring once the previous year's monthly burned-area files are complete; an interim release may follow an exceptional fire season. The physics is held fixed between refreshes — only the burned-area history, the regional level and updated inputs move. Older releases remain available; pin one, and re-derive rather than rescale anything baked from it. Details in README.md → *Versioning and refresh*.
v1.1-xgb
2026-05
never shipped. Listed as pending while the global model was retrained; superseded by the physics-based simulation-emulation line, shipped as v1.0, before release.
- Machine-learning wildfire likelihood. Gradient-boosted trees (XGBoost) trained on satellite active-fire detections with spatially-blocked cross-validation and climate-envelope-matched negatives (hold-out AUC 0.94). Inputs include peak-fire-season NDVI, canopy height, dry-month count, lightning density, categorical land cover, roads, and a climate envelope. Ships with a bootstrap-ensemble uncertainty layer and coverage.
Returns
v1.0
2026-09
First shipped release of the wildfire Returns line; companion to Wildfire Surfaces v1.0 and built from the same annual burn rate.
- Horizon burn probability at 90 m. rp50 / rp100 / rp250 / rp500: P = 1 − (1 + λ/k)^(−k·T), a negative-binomial annual burn count with mean λ (the Surfaces rate, including its local history anchor to the MODIS record) and dispersion k fitted per GFED basis region × land-cover group from the year-to-year variance of NASA MODIS MCD64CMQ burned fraction, 2001–2023. Poisson limit where the record is under-dispersed.
- Semantics. Non-burnable land reads 0 at every horizon; horizons are monotone; ocean and Antarctica nodata. Not an intensity product — flame length / fireline intensity at return periods is a later line.
- Known limits. Stationary rate over the horizon; no fuel recovery; regional dispersion applied at pixel scale. Inherits every Surfaces limit.
- Versioning and annual refresh. v1.0 is a conventional product version — the first release of this model line; the next refresh is v1.1, or v2.0 if the method changes. A Returns release always pairs with the Surfaces release of the same label. The label is not a year, so data currency is stated separately: burned-area record 2001–2023, global level 2014–2023, dispersion from the same record, WorldCover 2021 fuel through crosswalk v1.4, Ignition Layer v0.3. One refresh a year against the burned-area record, an interim release after an exceptional season; the physics is held fixed between refreshes and the dispersion is refitted from the extended record. Older releases remain available — pin one. Details in README.md → *Versioning and refresh*.
Other hazards
LandslideAvalancheEarthquakeWindstormExtreme heatExtreme coldFrost / freezeLightningSoil erosionHurricaneDroughtVolcanic EruptionTsunamiStorm SurgeHail