The 73% I cite publicly is the top-50 within 50m row (73.7% exact, rounded down).

10m gets noisy because many of these caves were recorded before consumer GPS got accurate. The model isn't really missing them, the dataset just doesn't know exactly where they are.

Other miss sources: bluff-face entries, canopy-shrouded entries, skid trails, agricultural micro-relief. Plus caves destroyed since they were recorded.

Methodology paper in prep.

cavefinder.app

#speleology #geomorphology

How Does CaveFinder work?

Tested against 13,000 known TAG caves with 1km square search areas, cave location placed randomly inside so the model doesn't learn it's always centered.

Within 50m: 40% top 10, 60% top 25, 73% top 50.
Within 25m: 23%, 32%, 38%.
Within 10m: 13%, 17%, 20%.

#caving #karst #LiDAR

Over 250 cavers are using it, many of whom are finding caves with it.

I'm giving a research talk at NSS Convention this year. Academic paper on the methodology coming within the year.

🔗 https://cavefinder.app
💬 Discord: https://discord.gg/9B5qaGARH
📘 FB group: https://www.facebook.com/groups/1314874967416327

The 27% it missed were mostly cave entrances on bluff faces, cliffsides, or under really dense canopy where ground-return LiDAR is sparse. Some are caves where the recorded coordinates are off. Hard limits of the data, not bugs in the app.

The site has ~5,000 publicly-known US cave locations for reference. Most cave locations are private and should stay that way. CaveFinder.app finds caves from public LiDAR data.