WLF · Machine learning
Wildfire
XGBoost · spatially blocked CV · envelope-matched negatives · hold-out AUC 0.94 · v1.1 · 2026-05
preview pending
Preview ships when the v1.1 global retrain finishes
FIRMS active-fire labels 2015–2023 (7.1M positives). Envelope-matched negatives restrict the training set to fire-capable climate envelopes (no Sahara/Greenland trivial negatives).
What ships
Surfaces — calibrated 0–1 probability per pixel, with bootstrap-ensemble uncertainty.
Every release: 0–1 score · 5-class band · per-pixel uncertainty · input-coverage mask · STAC item · methodology doc.
Inputs
NDVI peak-fire-season, canopy height, dry months, lightning density, landcover (categorical), roads, climate envelope
Validation
Spatially blocked hold-out AUC 0.94 — cross-validation blocks prevent leakage across nearby pixels.