“Compared to anything I’ve seen from Microsoft and Google, this dataset appears to have much more accurate building footprints to the point I initially thought they were manually surveyed.”
Oak Ridge ran a neural network over Maxar satellite imagery and labelled 131 million US buildings. Almost 62 million have no address and almost 68 million have no ZIP code. The validation field says Automated for 99.6 million, Unverified for 32.2 million and Manual for 2,731. A human checked about one building in 48,000.