# Data Sources — Petrel Karst & Depressions Layer v0.1

| Dataset | Role | Licence | Where verified |
|---|---|---|---|
| GEDTM30 v1.2 (OpenGeoHub) | Bare-earth elevation, the input to depression filling | CC BY 4.0 | Zenodo record; PeerJ 2025. Verified at source, not from a description column. |
| GLiM v1.0 (Hartmann & Moosdorf) | Lithology, reclassed to the solubility ordinal | CC BY (PANGAEA) | PANGAEA distribution |
| GSHHG-f (Wessel & Smith 1996) | Land/ocean mask, full resolution | Public domain / LGPL | Canonical Petrel land mask |
| Florida Subsidence Incident Reports (SIRS) — **Florida Department of Environmental Protection, Florida Geological Survey (DEP/FGS)** | **Validation only** — never an input to the bands | Public: *"no proprietary rights may be attached to it"*. Attribution required; narrow no-resale-to-government clause (see below) | SIRS Disclaimer read at source 2026-09-18, floridadep.gov/fgs/sinkholes/content/sirs-disclaimer — the terms below are quoted from it, not from a portal description field |

## Sources deliberately NOT used

| Dataset | Why not |
|---|---|
| FABDEM | Non-commercial terms. NC restrictions follow derivatives — a raster traced from an NC DEM inherits them, and "our derivative is a new work" is not a defence. Banned platform-wide. |
| WOKAM (global karst aquifer map) | Licence unresolved: "other restrictions", three non-German co-rightsholders. GLiM supplies the same lithological signal and is CC-BY. |
| BGS National Karst Database | Licence "Not set", enquiry-only. A commercial-use request is outstanding. |
| COST TU1202, Galve et al. (Spain), Lei et al. (China) | "Free academic" does not clear commercial use, and training a commercial model on them is a more direct use than a raster derivative, not a lesser one. |

## Credit where it is required

**The validation in this layer rests on data collected by the Florida Department
of Environmental Protection, Florida Geological Survey (DEP/FGS).** Their
disclaimer states: *"we expect the DEP/FGS to receive proper credit when you
utilize these data sets"*, and this section is that credit. Every performance
figure in `model_card.md` and `validation_report.md` is measured against their
Subsidence Incident Reports database.

## The Florida SIRS terms, read at source

Verified 2026-09-18 against the SIRS Disclaimer itself rather than a portal
metadata field. The operative sentence:

> *"Since part of this data was developed and collected with U.S. government or
> State of Florida funding, no proprietary rights may be attached to it in whole
> or in part, nor may it be sold to the U.S. government or the State of Florida
> as part of any procurement of products or services."*

Read plainly, that is an **anti-enclosure** clause — it exists to keep publicly
funded data public — plus one narrow bar on reselling *the data* into a
government procurement. It is not a restriction on using the data to measure
something.

**It does not reach this layer.** Not one SIRS record, coordinate or attribute
is redistributed here; the shipped bands are derived from GEDTM30 and GLiM. What
we publish is an aggregate statistic *about* our own raster, computed against
their inventory — the same relationship MTBS and NBAC have to Petrel's wildfire
product. The no-resale bar binds the inventory, not a product validated against
it.

⚠️ **A correction to our own earlier note.** Petrel's internal record previously
described this restriction as landing on the outbound product. That was written
when the inventory was a candidate **training** input, where the data would have
been embedded in fitted weights. For **validation-only** use it does not apply,
and treating it as though it did would have been a false licence alarm — the
GEBCO error, which nearly drove us off a public-domain source onto a banned one.

## DEP/FGS's own accuracy caveat, and why we honour it

The same disclaimer warns that the records *"have not been field checked or
verified by DEP/FGS professional geologists"* and that *"the locations of many of
these subsidence incidents have not been verified"*, and that DEP/FGS *"assumes
no responsibility for errors"*.

This is precisely why the validation uses only `TRUE_SINK` = **Y** (inspector
verified) and **N** (investigated, confirmed not a sinkhole) and **excludes the
4,232 unverified reports**. Our own measurement is the empirical form of their
caveat: the unverified reports score **AUC 0.529** against the confirmed
non-sinkholes — indistinguishable from chance. The source said its unverified
records were unreliable; the data agrees.

## Note on the elevation source

Bare earth matters here specifically. A surface DEM places buildings and canopy
in the terrain, and both manufacture closed depressions that the fill cannot
distinguish from dolines. Measured, the surface-model variant scores lower on
the same test set than the bare-earth one.


## Instruments — used to grade this layer, not to build it

These datasets never touch a shipped pixel. They are listed because a validation
number inherits every dependency of the instrument that produced it.

- **Maus, V. et al. (2022).** *An update on global mining land use.* Scientific
  Data 9(1), 433. Distributed via PANGAEA, https://doi.org/10.1594/PANGAEA.942325,
  licensed **CC-BY-SA-4.0**. 44,929 polygons covering 101,583 km² of mining land
  use, interpreted from 2019 Sentinel-2 imagery.

  Used to measure what share of this band's deep values are mining rather than
  karst — the answer being 0.24% above 10 m. CC BY-SA 4.0 §3(b) attaches only on
  *Sharing Adapted Material*, so measurement of our own raster against it carries
  no obligation and no restriction. It is deliberately **not** used to mask or
  otherwise compute any shipped value.