The platform

Built for scale,
engineered for decisions.

Petrel is the infrastructure layer beneath site selection, climate disclosure, and underwriting triage — three workflows where a comparable 0–1 score across hazards beats a single return-period curve. Six commitments make that possible.

01 — Resolution

90 m, every place on Earth

Every hazard, one grid. Native 90 m output in EPSG:4326, aligned across all 24 perils so portfolios score on the same coordinate frame. No mosaic seams, no national-boundary discontinuities, no ocean/land step changes outside the explicit coastline mask.

02 — Provenance

Auditable by the pixel

Every release ships a per-pixel uncertainty layer, a per-pixel input-coverage mask, a STAC item, and a versioned methodology document. Every input dataset is cited by source, version, and download date. When the methodology changes, the version number changes.

03 — Calibration

One scale, every peril

The same 0–1 calibration philosophy runs across all 24 perils — supervised ML where event inventories support it, Gumbel/GEV statistics for return-period perils, RUSLE for soil, site-amplified PGA for earthquake. The reason a portfolio cross-peril comparison actually means something.

04 — Distribution

Three ways in

Free public sample COGs for every live peril at data.petreldata.io — pull with curl, open in QGIS / rasterio. Hosted point-query API in private beta (Stripe-metered) for single addresses and portfolios. Full bulk vintage downloads — every peril, every release, as Cloud-Optimized GeoTIFFs — sold under enterprise contracts via sales.

05 — Latency

Annual vintage, versioned releases

Each Petrel Surfaces vintage is a frozen, citable snapshot. New vintages annually, point releases when methodology improves or input data refreshes. Old vintages remain accessible — once you score a portfolio against vintage 2026, that scoring is reproducible forever.

06 — Open methodology

No black boxes

Every model, every input, every training inventory, every validation cohort is documented publicly. The methodology PDF for each peril describes the algorithm, the data sources, the calibration philosophy, and the known limitations. The science is the product.

The technical stack

Inputs

09
  • Copernicus GLO-30 DEM (elevation, slope, aspect, TWI)
  • ESA WorldCover 10 m + Copernicus Global Land Cover
  • ECMWF ERA5 monthly extremes (1991–2020)
  • NOAA IBTrACS tropical cyclones
  • NASA FIRMS active fires + MODIS NDVI
  • GEM 2023 seismic hazard + USGS Vs30
  • WGLC lightning climatology
  • Potapov 2021 canopy height
  • OpenStreetMap roads + buildings

Methods

08
  • XGBoost (gradient-boosted trees) with spatial cross-validation
  • Gumbel and GEV extreme-value statistics for return-period perils
  • RUSLE empirical erosion model
  • Site-amplified PGA for earthquake
  • Envelope-matched negative sampling
  • Bootstrap ensembles for ML uncertainty
  • Closed-form delta-method standard error for extreme-value SE
  • rasterio.warp.reproject for elevation-lapse downscaling

Outputs

06
  • Surfaces: Cloud-Optimized GeoTIFF per hazard (calibrated 0–1 score)
  • 5-class risk classification (Very Low → Very High)
  • Per-pixel uncertainty raster (bootstrap σ for ML, delta-method SE for Gumbel/GEV)
  • Per-pixel input-coverage mask
  • STAC 1.0 catalog with full provenance
  • Versioned S3 buckets (s3://petrel-data/vintages/{year}/)