In practice, visible roads are often initially mapped using imagery alone, and later adjusted using GPS data if available.
It is widely recognized that GPS accuracy can be limited in areas with dense forest cover or mountainous terrains due to weak signal reception (drifts).
However, relying solely on imagery for alignment poses its own challenges, as it can have inherent offset issues that make it unreliable.
While I haven’t found official statistics, I believe that navigation and routing constitute the primary use-case for OpenStreetMap data. Considering this, I tend to prioritize, in mountainous and remote areas, quality averaged out GPS-traced geometry (e.g. Strava) over imagery alone.
Since satellite imagery is typically not visible to end-users in their applications, the limitations of GPS accuracy may go unnoticed. Conversely, if a road aligns with the imagery but deviates from the averaged GPS traces in an area with known accuracy issues, it may lead to complaints about poor map quality from end-users.
The main drawback of favoring GPS-traced geometry is that mappers may encounter conflicting road alignments that do not match the available imagery. However, it is important to acknowledge that imagery itself may not always represent the ground truth accurately due to offset issues.
We strive towards the most accurate roads alignment based on the real word, using multiple sources to achieve this is a good idea. However, we shouldn’t compensate alignment for inaccurate GPS, because when GPS gets improved over the decades we are left with inaccurate roads.
As said in the other topic, mapping requires delicate balance. The most accurate in my opinion is the Lidar. It is affected by the atmosphere when air is layered (stratified) where it does not move much. On that model then aerials are mapped onto to create orthophotos. There is often a shift, can be more than one meter. Then comes GPS
In our governmental GIS web height data derived from Lidar model can be queried by drawing lines. That helped me find a hole in a buried hose from pressure/height calculations. Astoundingly correct! No way to get that from consumer GPS.
Arial imagery can have offsets of several meters - right. Just look at different image sources at the same location (if not derived from the same raw image).
Single GPS traces of consumer devices are usually not better than 3 to 5 m, but often with heavy deviations of dozens of meters caused by poor satellite geometry, multipath effects etc.
Heatmaps generated out of several traces do some sort of averaging and thus reduce “side jumps”, but are still affected by “bad traces”.
Newer hardware using more GNSS systems are less prone to geometry problems and two-frequency-sensors have a better accuracy.
But most of the traces uploaded to OSM or Strava are up to now of the old sort.
Keeping this in mind you may carefully adjust image offsets. But these offsets change sometimes within a few dozen meters and you have reasonable traces typically only along main roads. Ignoring objects mapped by imagery and just correcting road geometry isn’t a good idea since then you will have e.g. streets running through buildings.
In some regions you may have exact maps/images legally usable by OSM from cadastre or orthophotos (LIDAR). This is definetly preferrable even over GPS traces, but precise and actual maps are very rare.
In short: Geometry correction is highly a matter of balancing and this balance may vary remarkably from scene to scene.