DETECT · PRIORITIZE · EXPLORE

Karst survey is no longer left to chance

KarstPro prioritizes your surveys with a morphometric score. A QGIS plugin that turns LiDAR into georeferenced field targets — sinkholes and shafts — reproducible, free of exploration bias. Available in France and, now, in Spain (experimental).

QGIS ≥ 3.34 · Python · PolyForm Noncommercial 1.0 · 🇫🇷 France · 🇪🇸 Spain (beta)

QGIS with the project generated by KarstPro: detected sinkholes, P1/P2 targets, geology, hydrology, topo network, hillshade and KarstPro tools in the Processing toolbox
Real QGIS — full project generated by KarstPro (Bayard-sur-Marne)

The problem

Days of GIS work per massif

National LiDAR now covers almost all of France at 1 point/m², but turning a raw point cloud into field targets stays manual and tedious.

01Slow

Download, cleaning, detection, vectorization — repeated by hand for every sector.

02Biased

Cavers return to areas they already know. Untouched sectors stay out of sight, not out of interest.

03Not reproducible

Two operators, two results. Comparing massifs to each other is nearly impossible.

The pipeline

Six automated steps, 5 minutes

From raw point cloud to a GeoPackage ready for QField. Public sources cross-referenced automatically — no account, no API key. The pipeline adapts to the chosen country : France (IGN, BRGM) or Spain (CNIG, IGME) — same output, same tools.

01 · DATA.GEOPF.FR / CNIG

LiDAR acquisition

COPC LAZ via IGN API (France) or PNOA 1 m tiles (Spain, 5 m WCS DTM fallback), automatic retry on network error.

02 · PDAL / GDAL

1 m DEM

Ground-class filtering (LAS 2), TIN interpolation → 1 m raster.

03 · WHITEBOXTOOLS

Sinkholes & shafts

Wang & Liu sink filling + D8 flow for sinkholes; vertical shaft detection by void signature (NODATA), filtered against buildings and water (BD Topo in France, SIOSE in Spain).

04 · RASTERIO / SHAPELY

Vectorization

Depth from DEM difference, 1 m threshold, 12 morphometric attributes, altitude auto-read from the DEM.

05 · BLOC A v2

Scoring

Score out of 100 (8 components), BRGM geology (France) or IGME (Spain), three-point dip estimate.

06 · GPKG / GPX / QFIELD

Field export

QField project, Garmin GPX, referenced water points, AI prompt with geocoding and dye-tracing injected.

P1 ≥ 55 — priority P2 35–54 — secondary P3 25–34 — moderate < 25 — excluded

Automatic prioritization by learned model, validated out-of-sample : Barrois applied by default (AUC 0.65–0.72 vs. 0.57 for manual weights), and four regional opt-in models — Jura plateau, Lot, Dordogne, Ardèche (AUC 0.60–0.64), France only. A multi-model router picks the domain from local geology ; outside a validated domain (or in Spain), it falls back to the manual score — never a silent model application. A "Diagnose a model" tool empirically checks, against your own known caves, whether a model fits a given commune — without ever activating it on its own. The score is still computed purely from LiDAR and geology, never from proximity to a known network : an untouched massif scores just as well as a well-known one.

Map of France showing the application zones of KarstPro's 5 learned models: Barrois (auto) and Jura plateau, Lot, Dordogne, Ardèche (opt-in), with nominal zones and veto radii
Validated domains — outside them, back to the morphological score

Field & AI

From desk to field, friction-free

  • Ready-to-use QField project — symbology by priority (P1 red, P2 orange), data-entry forms, built-in GPS.
  • GPX waypoints — compatible with Garmin, OruxMaps, CalTopo, with score and depth in the description.
  • AI survey plan — structured prompt (clusters, morphometry, dip, dye-tracing) to paste into Claude, ChatGPT or Gemini. Returns : prioritized targets by day, ordered GPX, Word report.
  • Field return → inventory — finds (caves, targets, visited shafts) are exported as CSV with photos and altitude, importable into Karst Entry, which deduplicates. PDF sheet per discovery.
  • Survey memory — re-preparing a study no longer loses field verdicts: targets with no interest are archived, already-visited or already-known ones are removed from the to-do.
  • Redo a study — replays an existing preparation with the same parameters (useful after a processing update), no re-entry, and without losing field verdicts already recorded.
QField on a phone: survey target map with contour lines
QField on a phone: navigating to a P2 target in the field

Limitations

A prioritization tool, not a decision-maker

Field expertise is irreplaceable. A P1 sinkhole can be a shallow depression with no continuation. KarstPro says where to look — the caver decides.
Flat entrances are undetectable. Below 40 cm of depth, no usable LiDAR signal. A physical limit, not an algorithmic one.
Calibration by domain. Five validated models — Barrois (auto), Jura plateau, Lot, Dordogne, Ardèche (opt-in). Elsewhere, the manual exploratory score applies — AUC caps at ~0.60–0.65, a physical limit of surface morphology.
Dense forest. Vegetation artifacts mask sub-metric sinkholes under heavy canopy.
Spain : exploratory pipeline. PNOA LiDAR, IGME geology, SIOSE filtering — but no learned model and no historical-subsidence signal yet. Manual score only, lightly tested in real conditions.