# Model Card — Petrel Ignition Density Layer v0.3 Global ML

Following the Model Card framework of Mitchell et al. (2019), "Model Cards
for Model Reporting", FAccT.

## Model details

- **Model name:** Petrel Ignition Density Layer v0.3 Global ML
- **Version:** v0.3-global-ml
- **Vintage code:** `ignition_v0.3_global_ml`
- **Release:** July 2026
- **Owner:** Petrel
- **Model type:** Composite of (a) a supervised gradient-boosted classifier
  (scikit-learn `HistGradientBoosting`) for the anthropic component, (b) an
  empirical per-Köppen conversion of observed flash climatology for the
  lightning component, and (c) a probabilistic-OR combination into a relative
  total propensity
- **Native resolution:** ~5 km (0.05°), global 60° S – 84° N — **not 90 m**
- **License:** CC-BY-NC-4.0
- **Citation:** see [`CITATION.cff`](CITATION.cff)
- **Contact:** layers@petreldata.io

## Intended use

### Primary intended uses

- Spatial ignition weighting for wildfire simulation (Cell2Fire ignition
  sampling — this Layer's role inside the Petrel Wildfire Hazard)
- Landscape- to regional-scale wildfire risk screening: where is ignition
  pressure human-dominated vs lightning-dominated
- Research in fire geography, human-fire interaction, and the
  wildland-urban interface
- Prioritization inputs for prescribed-burn planning and conservation risk
  assessment

### Out-of-scope uses

- **Absolute ignition-rate estimation from `total`.** The total layer is a
  relative propensity in [0,1], normalized within this vintage. It is not
  "ignitions per year" and must not be used as one. (The `lightning`
  sub-layer is also a relative index — it was mislabelled as a rate through
  v0.2 and corrected in v0.3.)
- **Parcel- or property-level assessment.** ~5 km cells aggregate entire
  towns and vegetation mosaics.
- **Real-time / incident-time operations.** Inputs are multi-year
  climatologies; nothing here updates at incident tempo.
- **Actuarial or regulatory use without independent validation.** v0.3 has
  not been quantitatively validated (see below).

## Training data

### Anthropic classifier

- **Labels:** NASA FIRMS active-fire detections (MODIS + VIIRS, 2015–2023),
  reduced to binary fire presence per cell. Note these are *fire detections*,
  not confirmed ignition points.
- **Features:** GHSL R2023A population, built-up surface, and settlement
  class (2020 epoch); HYDE 3.3 cropland and grazing area (2020 AD slice);
  Beck 2018 Köppen-Geiger class.
- **Sample:** up to 1.5 M valid cells of the 0.05° working grid, randomly
  sampled; early stopping on an internal 10% validation split.
- Labels are **globally distributed** (unlike region-restricted label sets),
  so no geographic extrapolation beyond the label domain is needed.

### Lightning conversion

- **Flash climatology:** NASA LIS/OTD 0.5° HRFC v2.3.2015 (combined
  OTD + TRMM-LIS record).
- **Calibration:** per-Köppen-class ratio of FIRMS lightning-attributable
  fire density to flash density, clipped to [0, 0.1] fires/flash; classes
  with <50 valid cells carry no estimate.

## Evaluation

**v0.3 has no published quantitative accuracy metrics.** Validation to date
is qualitative: agreement with GFED fire geography plus structural sanity
anchors (the wildland-urban-interface hump in the anthropic layer, tropical
continental lightning maxima, near-zero propensity over fuel-less deserts and
dense-urban cores). Quantitative skill scoring against GFED ignition
attribution is the flagged v1.0 work item. See
[`validation_report.md`](validation_report.md) for exactly what was and was
not checked.

## Uncertainty

**No per-pixel uncertainty layer ships in v0.3.** The anthropic component is
a single gradient-boosted model — no ensemble spread, no bootstrap interval —
and no post-hoc probability calibration step was applied beyond the model's
log-loss training. An uncertainty companion layer is a v1.0 candidate.

## Known biases & failure modes

| Failure mode | Surface | Status in v0.3 |
|---|---|---|
| **Agricultural-burning conflation** | FIRMS detects all vegetation fire; in cropping regions the anthropic layer partly reflects burning *practice*, not wildfire ignition risk | Documented limitation; no crop-fire filtering in v0.3 |
| **Detection bias in labels** | FIRMS misses small, short-lived, and cloud-obscured fires; label prevalence varies with overpass and cloud climatology | Documented; inherent to satellite active-fire labels |
| **Epoch mixing** | GHSL 2020 epoch, HYDE 2020 slice, FIRMS 2015–2023, LIS/OTD ~1995–2014 — treated as one contemporary climatology | Documented; acceptable for a climatological layer |
| **Per-Köppen constancy** | Lightning ignition efficiency is constant within a Köppen class; real efficiency varies with fuel moisture and storm type | v1.0 candidate: finer regionalization |
| **Vintage-relative total** | The p99 normalization means `total` values are not comparable across vintages | Documented; absolute calibration is the v1.0 item |
| **Köppen boundary artifacts** | Class-constant conversion factors can step at climate-zone boundaries | Visible in the lightning layer at some zone edges; smoothing considered for v0.3 |

## Ethical considerations

- **No PII.** All inputs are ≥1 km aggregate rasters describing settlement,
  land use, lightning, and fire detections. The product does not identify
  properties, parcels, or people.
- **Derived-product uncertainty.** Inputs (GHSL, HYDE, LIS/OTD, FIRMS) are
  themselves modelled or detection-limited products; v0.3 inherits and may
  compound their uncertainties.
- **Non-commercial license.** Commercial fire decisions should involve
  explicit human review, not automated reliance on an unvalidated v0.3
  research-grade layer — commercial licensing (licensing via https://petreldata.io/demo) exists
  partly to enforce that conversation.

## Caveats & recommendations

- Treat `total` as a *ranking* of where ignitions concentrate, not a rate.
- Cross-check the anthropic layer against regional ignition inventories
  (e.g. national fire databases) before relying on it in any one region.
- If you need absolute rates, use the `lightning` layer (physical) and wait
  for the v0.3 calibrated total.
- Report systematic biases to layers@petreldata.io — user feedback feeds the
  v1.0 validation 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