# Release notes — Petrel Karst & Depressions Layer

## v0.1 · 2026-09

First release. Global closed-depression depth and bedrock solubility at 90 m.

- **`karst_depression_depth_90m.tif`** — depth below the sink-filled bare-earth
  surface (GEDTM30, filled to each depression's lowest pour point). Zero where
  the pixel drains. Validated in Florida against 103 verified sinkholes and 108
  confirmed non-sinkholes: **AUC 0.649 [0.58, 0.71]**, i.e. 51% of verified
  sinkholes sit in a closed depression against 27% of confirmed non-sinkholes.
- **`karst_solubility_90m.tif`** — GLiM bedrock dissolution class (0 insoluble,
  1 possible, 2 carbonate, 3 evaporite). Shipped as context; it does **not**
  discriminate in Florida (AUC 0.520) because the surficial geology there is
  sand over carbonate.

**This layer deliberately ships no sinkhole probability.** The only sinkhole
inventory we can lawfully use commercially yields 103 verified positives, and
the 4,232 unverified reports in it were measured to carry no usable signal
(AUC 0.520 against confirmed non-sinkholes). A calibrated probability surface
needs a second hydrogeological setting first.

**A designed band was dropped on evidence.** Depression depth masked to soluble
bedrock was the original headline band; masking moves the Florida AUC from
0.649 to 0.520, so it was removed rather than shipped. The two bands ship
separately so users can combine them where they have regional evidence it
helps.

Depths are filled with a 256-pixel (~23 km) halo, a width measured across six
karst regions rather than chosen: at 64 px, South China cone karst is wrong by
up to 52 m, and without any halo a 41 m depression on a tile boundary reports
as flat.

Validation is one region. Treat the performance figure as indicative outside
Florida, and always quote it with its interval.