Reconciling OSM edits against a quarterly satellite road diff

Anyone running a basemap pipeline that pulls from both OpenStreetMap and a commercial imagery source eventually hits the same wall: two editors, OSM contributors and a satellite extraction run, touch the same stretch of road in the same quarter, and now you've got two versions of a segment that don't quite agree on geometry, name, or whether it exists at all. Neither source is wrong. They're just looking at the ground at different times through different methods.
Where the conflicts show up
The friction clusters in places both sources are actively watching: new subdivisions, highway interchange rebuilds, areas with active OSM mapping parties. Rural roads with one lane per county rarely cause trouble. Four patterns come up over and over.
Geometry drift. An OSM way was hand-traced from an old aerial layer or GPS traces and sits ten to twenty meters off the true centerline. The quarterly diff, pulled from 0.5–2 m imagery, places the same road correctly. Nothing's "wrong" in either dataset, but a naive merge either duplicates the segment or snaps one to the other and breaks connectivity on whichever way loses.
New construction timing mismatch. A subdivision road gets paved and mapped by a local contributor within days. Your imagery provider doesn't task a new capture for another ten weeks. By the time the diff arrives, OSM already has the segment, sometimes with a name, sometimes without. The diff shows it too. Now you're deciding which record owns the attributes.
Ghost segments. The diff flags a new paved surface that doesn't show up in OSM at all, usually because no local mapper has been through that area recently. This is the case the diff exists for, and it's also the easiest to handle wrong if your merge logic assumes OSM is always the newer source.
Decommissioned roads. A road closed or rerouted for construction still sits in OSM because nobody filed the deletion. The diff won't necessarily flag an absence as a change unless your extraction run is explicitly comparing against the prior quarter's layer, which is the whole point of running it as a diff rather than a fresh extraction each time.
A reconciliation workflow that holds up at scale
The pattern that works for production teams is a tiered pass that narrows down how much a human has to look at.
Start with a spatial join at a tolerance that matches your imagery's ground sample distance, not an arbitrary buffer. At 0.5–2 m GSD, anything the diff flags as a new or changed segment within a few meters of an existing OSM way is almost certainly the same road with a geometry or attribute discrepancy, not two separate roads. Anything beyond that tolerance with no OSM match nearby is a genuine candidate for a new segment.
Second, check OSM's edit history on anything that spatially matches. A way edited in the last 90 days by a contributor with a real edit history is probably current and should win on attributes like name and surface type, even if the diff's geometry is tighter. A way that hasn't been touched in two years and sits inside the diff's change footprint is a strong signal the diff is catching something OSM missed.
Third, route everything that doesn't resolve cleanly, duplicate geometries, conflicting surface tags, segments present in one source and absent in the other, into a short manual review queue instead of trying to force an automated rule to cover it. This is usually a small fraction of the total diff once the first two passes run, which is the point of doing them.
Fourth, decide your write-back policy before you start merging, not during. If your shop pushes corrections upstream to OSM, a clean workflow treats the quarterly diff as a changeset proposal reviewed by someone with local edit rights, not a bulk import. If you're maintaining a private routing graph, you can apply the merge directly, but keep the provenance tag on each segment (diff-sourced vs. OSM-sourced) so the next quarter's reconciliation knows what it's comparing against.
None of this requires rebuilding your network from scratch every quarter. It requires a diff that's actually a diff, change segments against the prior layer, not a full re-extraction you have to compare by hand. Road Extraction delivers that as a vector layer sized to merge into exactly this kind of workflow.
If your network's drifting out of sync with what's on the ground, a quarterly diff is worth a look.