# Validation Report — Petrel Ignition Density Layer v0.3 Global ML

## Validation status: qualitative only

**v0.3 has not been formally validated.** No quantitative skill score against
an independent reference has been computed for this vintage. What follows is
an honest record of the qualitative checks that were performed, the
structural sanity anchors the product passes, and the quantitative validation
scheduled for v1.0. Users who need demonstrated quantitative skill should
wait for v1.0 or validate v0.3 in their own region of interest before
relying on it.

## Qualitative checks performed

### 1. Agreement with GFED fire geography

The total-propensity field was compared visually against the Global Fire
Emissions Database (GFED) burned-area climatology (van der Werf et al. 2017).
The major features of global fire geography are present and in the right
places:

- the northern- and southern-hemisphere African savanna burning belts,
- the South American "arc of deforestation" and Cerrado,
- northern Australian tropical savanna,
- Southeast Asian dry-season burning regions,
- the boreal fire zones of Siberia and Canada (lightning-dominated, visible
  in the lightning sub-layer),
- low propensity over the Sahara core, Arabian interior, Atacama, and other
  fuel-less hyper-arid regions despite non-zero lightning or human presence.

This is a pattern check, not a skill score — see *What v0.3 does not
establish* below.

### 2. Structural sanity anchors

| Anchor | Expectation | Result |
|---|---|---|
| Wildland-urban-interface hump | Anthropic susceptibility ~0 in unpopulated wilderness, peaking at rural/exurban settlement densities, falling in dense-urban cores | Reproduced — the hump-shaped population response (Syphard 2007; Bistinas 2014) is visible in the anthropic layer; dense city cores read low despite maximal human presence |
| Tropical lightning maxima | Lightning layer maxima over the tropical continental convective centers (Congo Basin, northwestern South America, monsoon northern Australia) and near-zero at high latitudes | Reproduced — inherited from the LIS/OTD climatology by construction, modulated by per-Köppen ignition efficiency |
| Fuel-less lows | Both deserts (no fuel) and dense-urban cores (no fuel, despite people) read low in total propensity | Reproduced |
| Value ranges | `anthro` and `total` in [0,1]; `lightning` in [0,1] (relative index) | Confirmed at encoding time (uint16 cap + `PETREL_SCALE` tags) |

### 3. Automated product integrity (promote audit)

The Petrel promote audit machine-checks every release bundle: encoding tags
(`PETREL_SCALE`, coast-clip provenance), overview pyramids, required
documentation, and coastal reference points (e.g. Carolina Beach NC,
Malibu CA, Waikiki HI must resolve to data, not ocean). These are integrity
checks on the shipped rasters, not accuracy checks on the model.

## What v0.3 does NOT establish

- **No quantitative skill vs GFED** (or any independent reference). The
  planned benchmark — scoring the total-propensity ranking against GFED
  ignition/burned-area attribution, with per-biome breakdown — is the
  **flagged v1.0 work item**.
- **No absolute-rate validation.** `total` is a relative propensity in [0,1];
  it has deliberately not been compared against ignition counts. Absolute
  events km⁻² yr⁻¹ calibration is also a v1.0 item.
- **No regional ignition-inventory comparison** (e.g. national fire
  databases, USFS lightning-fire records) in v1.0.
- **No temporal validation** — the product is a static climatology; trend
  fidelity was not assessed.
- **No per-pixel uncertainty.** Single-model GBM without an ensemble; an
  uncertainty companion layer is a v1.0 candidate.

## Known limitations

1. **`total` is relative, not absolute.** The lightning source is normalized
   by its global 99th percentile before combination, so total values rank
   cells within this vintage only — they are not rates and are not
   comparable across vintages.
2. **~5 km native resolution.** Sub-cell heterogeneity (a town beside a
   forest, a highway through rangeland) is unresolved. Not for parcel-level
   use.
3. **FIRMS labels are detections, not ignitions.** Small, brief, and
   cloud-obscured fires are missed; in agricultural regions the anthropic
   layer partly reflects burning practice rather than wildfire ignition risk.
4. **Epoch mixing.** GHSL (2020), HYDE (2020 slice), FIRMS (2015–2023), and
   LIS/OTD (~1995–2014 observation record) are combined as one contemporary
   climatology; recent lightning trends are not captured.
5. **Per-Köppen lightning efficiency is class-constant.** Real flash-to-fire
   conversion varies within a climate class (fuel moisture, dry lightning);
   class-boundary steps can be visible in the lightning layer. Köppen classes
   with too few flash cells (<50) carry no lightning estimate and contribute
   zero to the total there.
6. **Coverage:** global land 60° S – 84° N. Ocean is NoData (GSHHG-f coast
   clip, `all_touched`); Antarctica is excluded by Petrel standard.

## v1.0 validation plan

1. Quantitative GFED benchmark: rank-correlation and per-biome skill of the
   total propensity against GFED fire attribution.
2. Absolute-rate calibration of `total` against ground-truth ignition
   databases (target unit: events km⁻² yr⁻¹).
3. Regional inventory spot-checks (CONUS lightning-fire records, European
   fire databases).
4. Per-pixel uncertainty companion (ensemble or bootstrap).

## How to validate locally

For your region of interest:

1. Compare the anthropic layer's spatial ranking against any local ignition
   inventory (points aggregated to ~5 km).
2. Check that known lightning-fire regions in your area appear in the
   lightning sub-layer.
3. If you use `total` as a sampling weight, verify results are insensitive
   to modest re-weighting of the two sources.
4. Report systematic biases to layers@petreldata.io — findings feed the v1.0
   plan.

## References

- van der Werf, G. R., et al. (2017). Global fire emissions estimates during
  1997–2016 (GFED4s). *Earth System Science Data*, 9, 697–720.
- Syphard, A. D., et al. (2007). Human influence on California fire regimes.
  *Ecological Applications*, 17(5), 1388–1402.
- Bistinas, I., et al. (2014). Causal relationships versus emergent patterns
  in the global controls of fire frequency. *Biogeosciences*, 11, 5087–5101.