On 27/03/2020 14:21, ael wrote:
Even then working from imagery can be very problematic. One of the Amazon mappers added a variety of roads to a construction site which were actually muddy tracks being used for the construction itself. A minority were destined to become highways or paths.
Co-incidentally, a similar problem occurred at https://www.openstreetmap.org/changeset/82602527 .
The area there is a yard for the storage of portable drilling rigs. What was added there as a service road (since deleted but was at https://overpass-turbo.eu/s/RZi ) isn't really one at all - it's just a convenient route through an area that happens to have no long vehicles on it. Different ages of imagery show similar but different routes over time as where particular long vehicles in the yard happen to be parked changes.
One other problem that shows up here is that "Esri World Imagery" seems to have been used as the source of the "service road". Esri's the most recent source here, but it is clearly offset compared to the others (and in particular to OS OpenData StreetView, which is pretty reliable for road centrelines).
Obviously this issue isn't due to using Facebook's "AI" at all, but it does suggest another question - how do we know what imagery layers were used for FB's suggestions in that area, what age they are, and how offset they are?
I usually end up adding data based on GPS position, helped by looking at OS OpenData StreetView (for alignment using road centrelines and historic waterway data), Bing (fairly old but not very much offset) and Esri (fairly new but often quite offset relative to anything else). If someone's trusting some other undocumented algorithm instead, how do they know whether it is accurate or not?
Best Regards, Andy _______________________________________________ Talk-GB mailing list Talk-GB@openstreetmap.org https://lists.openstreetmap.org/listinfo/talk-gb
