# Model Card — Petrel Fuel Layer v1.0 Global FM40 Crosswalk

Following the Model Card framework of Mitchell et al. (2019), "Model Cards for
Model Reporting", FAccT. Note: this Layer is a **deterministic crosswalk**, not a
trained statistical model — there are no learned parameters and no training set.
The "model" is a fixed rule table mapping ESA WorldCover land cover × Beck
Köppen-Geiger climate → Scott & Burgan FM40 class. The card is retained for
consistency, and because intended-use / limitation reporting is just as relevant
for a rule-based derived product.

## Model details

- **Name:** Petrel Fuel Layer v1.0 Global FM40 Crosswalk
- **Version:** v1.0-global-crosswalk
- **Vintage code:** `fuel_v1.0_global_crosswalk`
- **Release:** July 2026
- **Owner:** Petrel Data
- **Type:** Deterministic categorical crosswalk — Scott & Burgan (2005) FM40
  surface fuel-model class assigned by rule from two inputs (WorldCover 2021
  land cover × Beck Köppen-Geiger climate)
- **Native resolution:** ~270 m (9 arcsec), global 60° S – 75° N
- **Output:** one categorical uint8 raster (FM40 class code; 255 = NoData); no
  scale factor
- **License:** **CC-BY-4.0** (commercial use permitted)
- **Citation:** see [`CITATION.cff`](CITATION.cff)
- **Contact:** layers@petreldata.io

## Intended use

### Primary intended uses

- The surface-fuel grid for wildfire spread simulation (Cell2Fire — this
  Layer's role inside the Petrel Wildfire Hazard)
- Landscape- to regional-scale fuels mapping and wildland fire risk screening
- A globally consistent, commercial-clean FM40 base layer for fire-behavior
  research where LANDFIRE-derived products cannot be redistributed
- A land-cover-conditioned fuels reference for teaching and cartography

### Out-of-scope uses

- **Canopy / crown-fire modelling.** v1.0 is surface fuel *class* only; it ships
  no canopy cover, height, base height, or bulk density. Do not infer crown
  fuels from it.
- **Parcel- or stand-level fuel typing.** ~270 m cells and a ~1 km climate
  driver aggregate whole stands; use local high-resolution fuels data for
  site-level fire planning.
- **Reading the code as a magnitude.** The value is a categorical FM40 code —
  never average, scale, or bilinearly resample it.
- **Fuel-load or fire-behavior accuracy claims from the specific in-family
  code.** v1.0 is validated at the **family** level (NB/GR/GS/SH/TU/TL); the
  in-family code is a representative assignment.

## Inputs

- **ESA WorldCover 2021** (10 m land cover, CC-BY-4.0) → the land-cover class.
- **Beck et al. (2018) Köppen-Geiger** (~1 km present climate, CC-BY-4.0) → the
  biome context that refines family and in-family code.
- No labels, no learned covariates. See [`data_sources.md`](data_sources.md).
  US LANDFIRE FM40 is used **only** as an independent validation reference, never
  as an input — which is what keeps the derived layer CC-BY-clean.

## Evaluation

**v1.0 has no statistical accuracy score, by design** — a deterministic
crosswalk is validated by *plausibility*, not by held-out prediction error.
Validation to date:

- **Biome spot-checks** across the world's major fuel biomes (tropical forest,
  boreal taiga, Mediterranean shrub, temperate grassland, savanna, water bodies)
  — the assigned FM40 family matches the physically expected family.
- **US in-domain family agreement** — the crosswalk is compared against LANDFIRE
  FM40 ground truth at the family level (NB/GR/GS/SH/TU/TL) over CONUS test
  regions.

See [`validation_report.md`](validation_report.md) for exactly what was and was
not checked.

## Uncertainty

**No per-pixel uncertainty layer ships in v1.0.** The output is a hard
categorical assignment from a rule table; there is no probability, spread, or
confidence per cell. The dominant structural uncertainty is (a) the climate
proxy for open-vs-closed forest and (b) the ~1 km Köppen resolution at biome
boundaries. A confidence / ambiguity companion is a possible future item once
canopy height is added.

## Known biases & failure modes

| Failure mode | Surface | Status in v1.0 |
|---|---|---|
| **Humid-C savanna over-read** | Savanna/open tree cover in humid temperate (Köppen Cf/Cw/Cs) climates is assigned closed-canopy TL, lacking a canopy-height input to see it is open | Documented; the v1.1 canopy layer is the fix |
| **Coarse alpine / tundra transitions** | Köppen is ~1 km, so high-elevation and tundra fuel transitions are blocky | Documented; v1.1 adds elevation |
| **Class-only** | No canopy attributes ship; downstream crown-fire logic must not rely on this layer | By design; v1.1 item |
| **In-family code is representative** | The specific code within a family (e.g. GR2 vs GR3) is climate-rule-assigned, not measured fuel load | Documented; family is the validated unit |
| **WorldCover epoch** | Land cover is the 2021 epoch; recent disturbance (fire scars, clearing) after 2021 is not reflected | Documented; tracks the WorldCover release |
| **Köppen boundary steps** | In-family code can step at climate-zone edges | Visible at some zone boundaries; inherent to a climate-conditioned rule |
| **Coverage** | Ocean NoData; Antarctica excluded (GSHHG-f L1) | By standard |

## Ethical considerations

- **No PII.** All inputs are ≥10 m aggregate rasters (land cover, climate). The
  product identifies no properties, parcels, or people.
- **Derived-product uncertainty.** WorldCover and Köppen are themselves modelled
  / classified products; v1.0 inherits their errors and adds the crosswalk's own
  simplifications.
- **Commercial-clean by construction.** The whole point of v1.0 is a fuel layer
  that can be reused and redistributed commercially (CC-BY-4.0) without the
  CC-BY-NC / share-alike encumbrances of the ML line — but a permissive license
  is not a correctness guarantee. Operational fire decisions should involve
  local fuels data and human review, not automated reliance on a ~270 m global
  crosswalk.

## Caveats & recommendations

- Key downstream logic on the **family**, not the specific in-family code.
- Reproject with **nearest-neighbour** only — this is categorical data.
- For crown-fire or canopy work, wait for v1.1 (real per-pixel canopy) or bring
  your own canopy layer.
- Cross-check against a regional fuels map before relying on the crosswalk in
  any one landscape, especially humid-temperate woodland/savanna mosaics.
- Report systematic biases to layers@petreldata.io — user feedback feeds the
  v1.1 refinement plan.

## References

Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B.,
Spitzer, E., Raji, I. D., & Gebru, T. (2019). Model Cards for Model Reporting.
FAT* '19. https://doi.org/10.1145/3287560.3287596