Thanks for the additional comments. Some clarifications:

• We consider the following highway tags to be roads: motorway, motorway_link, trunk, trunk_link, primary, primary_link, secondary, secondary_link, tertiary, residential, road, unclassified, living_street. We don’t consider highway=service or highway=track, as these generally represent access to fields, parking lots, driveways, etc. See p.6 of the paper. Note that even in countries like Germany where the road network was complete in OSM many years ago, these “other” highways are still growing rapidly (see the figure above, which shows trends for “other paths”).

• The figure that ~23% of the world’s roads are in the US comes from our analysis, and thus depends on our estimated level of completeness. But it’s comparable to other sources. For example, IRF World Road Statistics suggests that 18.4% of the world’s roads are in the US. This 18-23% range seems reasonable when one considers the low density of most US urban development (meaning that each km of residential road in the US serves many fewer people), and the extensive network of rural roads.

It’s an important point that the length of the road network is a moving target, as new roads are built. We don’t attempt to capture this dynamic. However, I’d be very surprised if physical roads were increasing faster than additions to the OSM database.

Thanks for the thoughts. I have done a little mapping in OSM, mainly with my students in class projects, but our conclusions are based on the visual assessment and modeling described in the paper. I’m curious why you think the numbers are way off? Note that large countries such as the US (which alone accounts for nearly a quarter of the world’s roads) push the global average up, even if many other countries are much less complete than the global average of 83%.

Thanks again to all of you for these ideas about how to improve and extend our analysis. We did quite a bit of additional work to sample lower-density areas, and refined our process to fit S-shaped curves. We conclude the at the global level, OSM was ~83% complete as of January 2016. Rerunning the models with the April 2017 OSM data give an estimated completeness of ~89%. (These figures are for streets only, and don’t say anything about other features included in OSM.)

The final paper is now published, and available here. The code and supporting information is posted to GitHub. We look forward to others improving on our effort.

Thanks, Joost, for these ideas. I wanted to give a quick update. Based on all these comments, we are going to spend some more time doing more visual inspections. In particular, we are going to sample more systematically across the entire rural-urban gradient. We hope to have more to share in the next month or two.

Thanks for all the additional comments. That’s interesting to hear about Flanders and Chile, and the point about the lack of good satellite imagery remaining is well taken.

One question: which countries do you think are the best mapped (especially outside Europe)? What about countries like Haiti and Nepal, where HOT has done so much work? What others?

A few clarifications:

• We use the Google hybrid OpenLayers plugin within QGIS for the imagery. The Google Maps layer provides some additional information, but we used the imagery as the primary source, as the Google data may not be complete either. So the resolution for the visual assessment is whatever Google offers.
• each grid cell is about 2.5 square km, although due to the projection used it varies with latitude
• we do exactly what joost suggests - we look at the shape of the growth curve of the road network, and crosscheck with the visual inspection.
• we model imports through allowing for jumps in the growth curve. You can see the jumps in the predictions here. Saudi Arabia is a nice case in point.

Thanks for the comments. The point about saturation in urban roads is a good one.

We do cross-check, as Mikel says, against a visual assessment using a small sample (20 grid cells per country) of satellite imagery. For now, we only use the estimated completeness from the S-shaped curve (i.e., its asymptote) where it agrees with the visual assessment (i.e., it falls within the 95% confidence interval). This should mitigate most cases where growth in OSM road length has saturated but the network is still far from complete.