This dataset provides Microsoft's ML road detections in Parquet format, including 54.2 million kilometers of roads detected across 235 countries worldwide.
A cloud-native mirror of the road detections Microsoft mines from Bing Maps aerial imagery.
Microsoft publishes this data at microsoft/RoadDetections as nineteen zipped TSV files of GeoJSON, one per UN subregion, totalling about 67 GB unpacked. That format is fine for a bulk download and useless for a query. This catalog carries the same roads as GeoParquet partitioned by country, so a reader can pull one country, one bounding box, or the whole planet without unpacking anything.
One collection:
This is a mirror, not the source. Microsoft produces the data. This catalog reformats and hosts it, and adds nothing to it. Geometries, widths, and country assignments are exactly as Microsoft published them.
Upstream drop: 2025.04.28, the current release as of 2026-08-14. Roads were converted to
GeoParquet in October 2025. If Microsoft publishes a newer drop, this mirror will lag it until it is
re-synced; the updated field on the catalog and the collection records the last sync.
For what the detections are, how they were produced, what they do not cover, and the traps in the country codes, read the collection README.
ODbL-1.0, the Open Data Commons Open Database License.
Microsoft's LICENSE file reads in full:
Data in this repository has been licensed by Microsoft under the Open Data Commons Open Database License (ODbL).
ODbL is share-alike. A Derivative Database you publish has to be offered under ODbL as well, and this mirror is. Microsoft states no required attribution wording, so ODbL section 4.3 obliges you to supply one. This works:
Contains road detections from Microsoft, mined from Bing Maps imagery, licensed under ODbL.
Microsoft's terms do not grant rights to its names, logos, or trademarks, which is why no Microsoft branding appears anywhere in this catalog.
Every partition, read as one table:
One country, which reads one file instead of 235:
Agents should start at AGENTS.md, which carries the tested query recipes, the join keys, and the quirks that otherwise produce confident wrong answers.
Catalog metadata lives in nlebovits/microsoft-ml-road-detections and is validated on every pull request. Corrections are welcome as issues or pull requests. Problems with the road data itself belong upstream, at microsoft/RoadDetections.
Contact: nissim.lebovits@radiant.earth