A GPS distance to the green is only as good as the map behind it. And that is exactly the problem: most golf courses in the world are not fully mapped in any open map. Of 897 golf facilities in Germany that OpenStreetMap knew in October 2026, 335 were mapped well enough for GPS. 318 did not have a single hole drawn in.
We didn't want to trace them by hand, course by course. So we taught an AI to read golf courses from above. The result powers the DeepSwing golf course GPS. Here is how it works and how good it really is.
The teacher: courses that are already mapped well
Every model is only as good as its examples. Ours are the courses that are already fully mapped in OpenStreetMap: fairways, greens, tees, bunkers and water as clean areas. Those areas become training masks. They are paired with aerial imagery from open official sources: state orthophotos in Europe and NAIP in the US, each at about 0.3 metres per pixel.
That way the model learns what a green looks like not only in North Rhine-Westphalia but also in California or Denmark.
The model: five classes, pixel by pixel
The course AI is a segmentation model. It assigns every pixel a class: fairway, tee, green, bunker, water or background. We measured it on validation courses the model never saw during training:
| Class | found | precision |
|---|---|---|
| Green | 99.1% | 95.7% |
| Bunker | 94.1% | 95.2% |
| Fairway | 93.7% | 91.4% |
| Tee | 81.2% | 62.0% |
| Water | 60.4% | 74.1% |
"Found" means: of all real greens, the model detected 99.1 percent. "Precision" means: of everything the model takes for a green, 95.7 percent really is one. The centre of a detected green is a median 0.6 metres from the real one, and less than 1.4 metres in 90 percent of cases. For a GPS distance that is more than enough.
To be honest: tees and water are the weak classes. Tees are small and often look like ordinary grass from above. Ponds hide under trees or reflect the sky. That's why we don't rely on segmentation alone for them.
From pixel to hole
Areas alone are not a golf course yet. Which tee plays to which green? In which order? With what par? A second, learned model decides that.
It scores every possible tee-green pair between 70 and 650 metres apart: is there fairway on the line between them? Is another green in the way? Does the length fit a hole? A classic optimisation method solves the best assignment of all pairs. The order follows from the shortest walks from each green to the next tee, the par from the length.
A classifier also separates hole greens from practice and chipping greens. We counted what such rules are worth: a detected green only counts if a fairway of at least 80 metres lies within 25 metres. With this rule, 88 percent of greens on 15 reviewed courses were real hole greens; without it, only 55 percent.
Checked before anything goes live
No AI proposal lands in the app unchecked. A course must have the right number of holes, no crossing lines of play, plausible walks from green to next tee and consistent pars. On top of that, every region must reach at least 90 percent correct holes in a sample review. Every course gets a quality grade from A to D; only A and B count as GPS-ready.
Terrain: laser scans instead of satellites
To know how far a shot really plays, you need heights. Here we use laser-scan elevation models; in Germany, the state survey's DGM1 with a one-metre grid. Measured against the official original, our heights are within 0.01 metres. A global satellite elevation model came in at 2.05 metres and was up to 7.9 metres off on one hole, because it measures tree tops instead of the ground.
From the heights we compute the plays-like distance: distance plus elevation change, divided by a factor for thinner air, about one percent per 290 metres above sea level. Hole 4 at Oberstaufen-Steibis is 458 metres long but drops 26 metres and sits at 926 metres altitude. It plays like 419 metres. More in Plays-like distance: how much elevation really changes.
The same heights make greens readable: for every green we compute height and slope on a one-metre grid. How to use the heatmap is in Reading greens with a heatmap.
What ends up on the phone
All that reaches the phone is the result: a course package of about 30 to 210 kilobytes per course. From it the app draws the map, computes distances to the green and hazards and builds the 3D terrain. Because the map is drawn from the data itself, it also works without a network.
Where we stand
As of October 2026 we have captured 40,057 golf courses worldwide, 10,229 of them GPS-ready with holes, greens and terrain. The pipeline runs region by region across 512 areas of the world.
And the pattern behind it, open data as a teacher, aerial segmentation and laser-scanned terrain, works for more than golf. More on the technology on our course AI page.
Frequently asked questions
How accurately does the course AI detect a green?
On courses the model has never seen, it finds 99.1 percent of greens at 95.7 percent precision. The centre of the green is a median 0.6 metres from the real one.
Where does the aerial imagery come from?
From open official sources: state orthophotos in Europe and NAIP in the US, at about 0.3 metres per pixel.
Why not just use OpenStreetMap?
Because many courses there are incomplete. In Germany, only 335 of 897 facilities were mapped well enough for GPS in October 2026. Where OpenStreetMap is good, we use it, including as a teacher for the model.
Can the method be used for other surfaces?
Yes. Open data as a teacher, segmentation of aerial imagery and laser-scanned terrain carry over to other surfaces.
If you need to recognise surfaces from the air, talk to us. And if you play golf: the DeepSwing golf course GPS, free on iPhone.