Course AI: mapping golf courses from aerial imagery

We taught an AI to read golf courses from above.

Most golf courses in the world are not fully mapped anywhere. An outline, maybe a few greens, nothing more. Our course AI closes that gap: a segmentation model trained on aerial imagery at about 0.3 metres resolution draws the playing surfaces. A second, learned model assembles the holes. And laser-scan elevation models provide the terrain down to centimetres.

Fig. 05 — Course AI & cartography
golf courses captured worldwide
40,057
of greens found on unseen test courses
99%
median error of the green centre
0.6m
terrain RMSE against official DGM1
0.01m
Aerial image of Golf-Club Bad Pyrmont
AI segmentation: fairways, greens, tees, bunkers, water
Fig. 05.1 — Golf-Club Bad Pyrmont, a validation course the model never saw in training. Aerial image: Geobasis NRW (dl-de/zero-2-0).

05.1 — Geodata & segmentation

Course AI & cartography

01

Learned from the best

The teachers are courses that are already fully and cleanly mapped in OpenStreetMap. Their areas become training masks. The aerial images come from open official orthophotos in Europe and NAIP in the US. That way the model learns what a green looks like, in North Rhine-Westphalia as well as in California.

02

From pixel to hole

The segmentation model produces areas. Which tee belongs to which green, in which order the holes are played and what par they have is decided by a second model. It scores every tee-green pair, solves the assignment optimally and fixes the round with the shortest walks. A classifier separates hole greens from practice greens.

03

Quality by design

AI proposals only go live when they pass strict checks: the right number of holes, no crossing lines of play, plausible walks and pars. Every region must reach at least 90 percent correct holes in a sample review. Every course gets a quality grade from A to D.

04

Cartography in centimetres

Terrain comes from laser-scan elevation models with grids down to 1 metre, in Germany from the state survey's DGM1. Against the official model we measure 0.01 metres RMSE. From it we compute each hole's elevation profile, green slopes on a 1-metre grid and plays-like distance. In Bavaria we even measure trees crown by crown from the surface model.

05.2 — Capabilities

What it does.

In use in our products, every day, on real devices.

  1. 01

    Playing areas from aerial images

    Fairways, greens, tees, bunkers and water, drawn automatically.

  2. 02

    Holes assembled

    Which tee plays to which green, in which order, with what par.

  3. 03

    Green distances

    Front, centre and back for every green.

  4. 04

    Hazards with reach and carry

    For every hazard on the line of play.

  5. 05

    Plays-like

    An elevation profile every 5 metres and the distance a shot really plays.

  6. 06

    Green slope

    Slope on a 1-metre grid, the basis of the green heatmaps.

  7. 07

    Trees crown by crown

    Measured from the surface model in Bavaria, for real 3D courses.

  8. 08

    Worldwide

    40,057 facilities captured, 12,000+ playable in DeepSwing.

Golf Oberstaufen-Steibis
Fig. 05.2 — Golf Oberstaufen-Steibis: 18 holes, contours every 4 m from DGM1 laser-scan data.

05.3 — How it works

How it works

  1. 01

    Aerial image

    Open orthophotos at about 0.3 m resolution.

  2. 02

    Segmentation

    Fairway, green, tee, bunker, water.

  3. 03

    Holes

    Learned model: tee to green, order, par.

  4. 04

    Terrain

    LiDAR heights, plays-like, quality grade.

05.5 — FAQ

Frequently asked questions

Where does the course AI's data come from?

Only from open sources: OpenStreetMap (ODbL), official orthophotos and elevation models of the states, and NAIP in the US. Every source is kept with its licence and attribution.

How many golf courses does the course AI cover?

As of October 2026, 40,057 golf courses are captured worldwide. 10,229 of them are GPS-ready with holes, greens and terrain. The number grows with every run.

How well does the model detect greens and bunkers?

On validation courses the model never saw in training, it finds 99 percent of greens at 96 percent precision. The green centre is a median 0.6 metres from the real one. 94 percent of bunkers are found.

Can the course AI be applied to other areas?

Yes. The pattern of open labels as a teacher, aerial segmentation and laser-scanned terrain also works for sports facilities, agriculture, solar parks or infrastructure.

Contact

What should your camera recognise?

Object detection, body tracking, segmentation or cartography, on your device, in real time. Let's talk about it.