# Validation Report — Petrel Aridity Index Layer v0.1 Global ML

## Validation status: qualitative only

**v0.1 has not been formally validated.** No quantitative cell-wise skill score
against an independent aridity reference has been computed for this vintage.
What follows is an honest record of the qualitative checks performed, the
structural sanity anchors the product passes, and the quantitative validation
scheduled for v0.2. Users who need demonstrated quantitative agreement should
wait for v0.2 or validate v0.1 in their own region of interest first.

## Qualitative checks performed

### 1. Agreement with reference aridity geography

The continuous index and the UNEP class map were compared visually against the
CGIAR-CSI Global Aridity Index (Zomer et al. 2022) and the UNEP World Atlas of
Desertification. The major features of global aridity geography are present and
in the right places:

- the Sahara, Arabian, Iranian, Thar, Taklamakan/Gobi, Kalahari/Namib, Atacama,
  and Australian-interior hyper-arid / arid cores read near zero;
- the Sahel, Mediterranean rim, US High Plains, Central Asian steppe, and the
  Horn of Africa show the expected semi-arid gradients;
- the Amazon, Congo, Maritime Continent, and monsoon Asia read strongly humid
  (P/PET > 1);
- the arid/humid transitions track the reference maps' isolines.

This is a pattern check, not a skill score — see *What v0.1 does not establish*.

### 2. Structural sanity anchors (promote audit test points)

| Anchor | Expectation | Result |
|---|---|---|
| Sahara / Libya | hyper-arid, AI < 0.15 | 0.006 ✓ |
| Rub al Khali | hyper-arid, AI < 0.15 | 0.008 ✓ |
| Atacama | hyper-arid, AI < 0.20 | 0.041 ✓ |
| Singapore | humid, AI > 0.65 | 1.26 ✓ |
| Amazon | humid, AI > 1.0 | 1.69 ✓ |
| Carolina Beach NC | coastal land resolves to data (not ocean) | data ✓ |
| Antarctica | NoData (GSHHG-f L1 exclusion) | NoData ✓ |
| mid-Pacific | NoData (open ocean) | NoData ✓ |

The UNEP class layer resolves Sahara → class 1 (hyper-arid) and Amazon →
class 5 (humid), consistent with the continuous index.

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

The Petrel promote audit machine-checks every release bundle: encoding tags
(`PETREL_SCALE`, `PETREL_COAST_CLIP` provenance), overview pyramids, required
documentation, and the test points above (orientation + mask correctness).
These are integrity checks on the shipped rasters, not accuracy checks on the
underlying climatology.

## What v0.1 does NOT establish

- **No quantitative skill vs CGIAR-CSI** (or any independent reference). The
  planned benchmark — cell-wise correlation and UNEP-class confusion matrix
  against CGIAR-CSI Global Aridity Index, with per-continent breakdown — is the
  **flagged v0.2 work item**.
- **No PET-method validation.** The dominant uncertainty is the Thornthwaite
  (temperature-only) PET approximation; v0.1 does not quantify the disagreement
  vs Penman-Monteith. A per-pixel Thornthwaite-vs-PM difference layer is a v0.2
  candidate.
- **No absolute water-deficit validation** against flux towers or basin water
  balances.
- **No temporal validation** — the product is a static 85-year climatology;
  trend fidelity was not assessed.

## Known limitations

1. **Thornthwaite PET.** Temperature-only; biases the absolute index (over-PET
   in humid warm climates, under-PET in arid). Read the index as a relative
   ranking; the UNEP thresholds are the community-calibrated interpretation.
2. **~28 km native resolution.** Sub-cell heterogeneity (an oasis, a coastal
   strip, an orographic gradient) is unresolved. Not for local use.
3. **Polar PET blow-up.** Cold high-latitude cells with PET → 0 give very large
   P/PET; capped at 4.0 for encoding (already class 5, humid).
4. **CI is sampling-only.** The bootstrap CI reflects interannual precipitation
   variability, not PET-method or reanalysis structural error — so a small CI
   does **not** mean the absolute value is accurate, only that the 85-year mean
   is stable.
5. **Coverage:** global land 60° S – 84° N. Ocean is NoData (GSHHG-f coast
   clip, `all_touched`); Antarctica is excluded by Petrel standard.

## v0.2 validation plan

1. Quantitative CGIAR-CSI benchmark: cell-wise correlation + UNEP-class
   confusion matrix, per continent.
2. Penman-Monteith PET variant (1991–2024) and a Thornthwaite-vs-PM
   disagreement layer.
3. Regional spot-checks against national aridity / agro-climatic zonations.
4. Structural-uncertainty accounting beyond the sampling CI.

## How to validate locally

For your region of interest:

1. Overlay the UNEP class map against any national agro-climatic or
   desertification zonation and check the class boundaries agree.
2. Compare the continuous index against a local Penman-Monteith aridity product
   if you have one — expect systematic Thornthwaite offset, consistent sign.
3. Use `aridity_index_ci` to see where the long-term mean is well constrained.
4. Report systematic biases to layers@petreldata.io — findings feed the v0.2
   plan.

## References

- Zomer, R. J., Xu, J., & Trabucco, A. (2022). Version 3 of the Global Aridity
  Index and Potential Evapotranspiration Database. *Scientific Data*, 9, 409.
  https://doi.org/10.1038/s41597-022-01493-1
- Middleton, N., & Thomas, D. (1997). *World Atlas of Desertification* (2nd
  ed.). UNEP / Arnold.
- Thornthwaite, C. W. (1948). An approach toward a rational classification of
  climate. *Geographical Review*, 38(1), 55–94.
- Hersbach, H., et al. (2020). The ERA5 global reanalysis. *QJRMS*, 146(730),
  1999–2049.