I'm a little confused as to why Grab would hire a firm to update OSM via satellite images. OSM had to be more accurate than the outdated sat. images so why was there a need in the first place?
As far as the imagery, DigitalGlobe provides relatively up to date imagery for most of the planet. The errors are probably more from overinterpreting the satellite data than from it being obsolete.
Shameless plug: At tensorflight.com we have deep learning models that can do just that.
Even 2 months old can be outdated (e.g. can be missing houses built since, or showing houses that no longer exist). In some areas, this can be a very significant effect.
My point was that OSM data can be much more accurate than the imagery. Or not, depending on location.
I'm sure you can use the newest/best data for your company because you can pass on the costs to your clients.
In the "high spatial resolution" - i.e. 50 cm per pixel or less the well-known providers everyone is using are Digital Globe, Airbus or Nearmap. Above 50 cm it starts to get hard to confidently see enough detail.
Google maps terms of service actually prohibits using satellite imagery for feature extraction ( https://cloud.google.com/maps-platform/terms/ ).
Bing is the best on the cheap end. Actually, Bing donated imagery for OSM mapping purposes ( https://wiki.openstreetmap.org/wiki/Bing_Maps#Bing_Aerial_Im... ).
But cleaning a bunch of GPS tracks can be a real mess. Many GPS tracks don't include the error/accuracy measurements and just pretend to be infinitely-small points scattered in the general vicinity of what might be a road. I think people tend to overestimate how accurate their satnav or phone's GPS is, because in most uses it's "locked" to a road or other map feature.
Anyway, not an intractable problem by any means, but preprocessing and cleaning GPS data to derive the actual terrain can be quite difficult. Paying somebody to click repeatedly on a satellite image might be temptingly easier.
I wonder if someone couldn't do "deltas" of satellite imagery to produce some sort of heatmap of how quickly various areas are changing. That would give you an idea of what areas are 'safe' to map based on satellite imagery, and what areas are changing too quickly and need to be mapped some other way.
It would also give you a way of automatically flagging stale data on the vector map; if something hasn't been updated in 12 months and you know from imagery that the area is under heavy development, those map features are probably not trustworthy anymore.
The idea was to take satellite imagery of the same location from different timestamps and I believe use deep learning to determine if anything in the imagery changed.
I don't think they've released any code for that project yet though.
But for determining whether or not the imagery could be used in a certain region, it might make more sense to just compare the timestamps of the vector data and the imagery. For instance, if the road doesn't appear in the imagery, but the vector data timestamp is after the imagery timestamp, then keep the road. And don't add a road if it was covered by a map segment that was deleted some time between the imagery timestamp and the current time.