Norwegian Microsoft ML Buildings - More harm than good?

Discussed in https://github.com/orgs/OvertureMaps/discussions/455

^{Originally posted by Johannes-Andersen January 27, 2026}
Hey!

Has there been any analysis of false positives, and are there any benefits to using GlobalMLBuildingFootprints in Norway? Given that Norway is a coastal country, it seems like false positives around docks are even more prominent.

Also, based on a few random samples, the data appears to be from around 2015, which means many footprints refer to buildings that have since been demolished. Is newer data available? If not, I would have expected OSM editors to have mapped most buildings by that time making the data even less reliable.

I’d be interested to hear whether it’s worth keeping this dataset enabled. Flying around my local area in Rapid Editor and reviewing the ML building recommendations seems to yield very few valid suggestions. That said, this could be confirmation bias, and it might be more valuable in other parts of the country.

My main issue with the dataset is that there doesn’t seem to be a way to remove features or provide feedback so that bad, non-existent buildings get filtered out over time.

Boat as buildings - example 1.

Boats as buildings - example 2.

Out of date - example 3.

Thanks,
Johannes