# Petrel Karst & Depressions Layer v0.1

Global closed-depression depth and bedrock solubility at 3 arcsec (~90 m),
EPSG:4326, land only (Antarctica excluded by standard).

## Read this first: what this layer is, and what it is not

**This is not a sinkhole probability, and it must not be used as one.** It is
two measured surfaces that a sinkhole assessment would want as inputs.

We are not aware of any global calibrated sinkhole *hazard* model. That is the
opportunity, and it is also why this layer is scoped the way it is: the honest
thing to ship first is the evidence, not the first unvalidated probability
surface with a global footprint.

The reason is the training data, not the method. The only sinkhole inventory we
can lawfully use commercially is the Florida DEP Subsidence Incident Reports.
It holds 4,449 records, of which **103 are verified sinkholes** (`TRUE_SINK` =
`Y`); 108 are confirmed *not* sinkholes and 4,232 are unverified reports. We
measured whether those 4,232 could serve as weak positives: they cannot. They
score AUC 0.520 against the confirmed non-sinkholes — indistinguishable from
chance. A model fitted on ~200 labelled points across thirteen features would
fit noise and then validate on itself.

## Bands

| File | Type | Units | Meaning |
|---|---|---|---|
| `karst_depression_depth_90m.tif` | uint16 (scaled) | metres | Depth below the sink-filled bare-earth surface. 0 where the pixel drains. |
| `karst_solubility_90m.tif` | uint8 | class | 0 insoluble, 1 possible, 2 carbonate, 3 evaporite; 255 nodata |

`physical = raw * SCALE`, with `SCALE` in the band tags. Read the tag; do not
hard-code it.

## What the depression band is worth

Measured against the Florida inventory — 103 verified sinkholes vs 108
confirmed non-sinkholes, every point sampled from the shipped global raster:

**AUC 0.649, 95% CI [0.58, 0.71].** In plainer terms, **51% of verified
sinkholes sit in a closed depression against 24% of confirmed non-sinkholes** —
a 2x enrichment. Real, and modest. Treat it as a screening signal that
concentrates where to look, not as a probability that a given parcel will
collapse.

## What `karst_solubility` is worth

It is a lithological fact, shipped as context. It is **not** validated as a
discriminator: on the same Florida sample it scores AUC 0.520 [0.46, 0.58],
i.e. no better than chance. That is not a defect in GLiM — Florida's surficial
geology is sand *over* carbonate, so the mapped bedrock is the sand. It means
you should not use solubility alone to rank Florida-like terrain, and you
should expect it to behave differently where karst crops out.

**Do not mask depression depth by solubility** unless you have evidence from
your own region that it helps. We tried exactly that as a headline band and
dropped it: masking to solubility >= 2 takes the signal from 0.649 to **0.520
[0.48, 0.56]**, i.e. destroys it, for the reason above.

## Known limitations

- **Validation is one region.** Florida, one hydrogeological setting on one
  peninsula. Nothing here establishes performance in evaporite karst, cone
  karst, or mantled karst elsewhere. Treat the AUC as indicative, not global.
- **Depth is a terrain measurement, not a subsurface one.** A closed depression
  may be a doline, a kettle hole, a quarry, a playa or a DEM artefact. The
  layer cannot tell them apart, and the non-doline features are frequently the
  *deepest* ones — every major open-pit mine sampled reads over 100 m, and
  endorheic basins such as the Tibetan Plateau and Bodele average 20-50 m
  against Florida karst's 0.99 m. model_card.md quantifies all three classes
  and gives the reading rules.
- **Bare earth is the input** (GEDTM30), because buildings and canopy in a
  surface DEM manufacture closed depressions. Residual canopy or building
  bias in the source DEM propagates here.
- **A depression that is actively subsiding looks identical to one that
  stabilised ten thousand years ago.** There is no temporal component.
- **No groundwater term.** Water-table depth and its fluctuation are primary
  sinkhole drivers and no commercially-clean global grid exists; this layer
  simply does not represent them.

## Licence

CC-BY-NC-4.0. Commercial use requires a Petrel licence — see LICENSE.md.

## Files

`README.md` (this file), `LICENSE.md`, `methodology.md`, `model_card.md`,
`data_sources.md`, `validation_report.md`, `RELEASE_NOTES.md`, `CITATION.cff`,
`API_USAGE.md`, `bulk_download_instructions.md`, the STAC item, and the band
GeoTIFFs.