It's machine-learning generated "slop" honestly.
Looking at where I live and where I grew up the building heights are quite badly estimated.
- Some groups of houses around here that are more or less identically built but on sloped terrain are reported to have widly differing heights
- My neighbour building is reported to be half the height of this building (they're more or less equally high at 5 stories)
- A small office shack behind the neighbour building is reported to be taller than it (it's a single-story building, the neighbour building is 5 stories)
- The freestanding buildings on the farm where I grew up are like you said, badly combined, much of the estimation there seems to be dependent on shadows,etc.
Still useful for a sense of building density. If almost everything called out is some sort of construction, then the density map of the world is a realistic estimate of human occupation.
Or, they subtracted a digital elevation model from a digital surface model, ran a point-in-polygon match against an existing building dataset, and labelled the difference as the height of the building. No ML needed.