hotosm/vhr-highway-segmentation

VHR Highway Segmentation

Road centreline training data at very high resolution. Aerial imagery tiles from OpenAerialMap paired with OpenStreetMap highway geometries, covering every validated OAM road-mapping project on the HOT Tasking Manager from the last five years.

The building counterpart is hotosm/vhr-building-segmentation. Both index tiles by tile_id, so they can be joined for multi-task training.

Usage

from datasets import load_dataset

ds = load_dataset("hotosm/vhr-highway-segmentation", split="train")

Structure

One row per image tile.

Column	Description
image	256x256 RGB aerial chip, georeferenced
mask	Empty for this dataset, see Labels
label_geojson	Road centrelines for the tile, EPSG:4326
class_counts	JSON of highway values in the tile, e.g. {"residential": 3, "path": 1}
num_features	Road count in the tile, 0 for a negative sample
tile_id, tile_x, tile_y, tile_z	XYZ tile identity
bbox_west, bbox_south, bbox_east, bbox_north	Tile bounds, EPSG:4326
project_id, project_url, project_name, country, organisation	Source Tasking Manager project
imagery_url	OpenAerialMap TMS endpoint the chip came from
feature_type	OSM tag key the labels were filtered on, always highway here

Tiles are zoom 19, so a pixel covers roughly 0.25 m and a tile spans about 63 m, varying with latitude.

Labels

Labels come from OpenStreetMap through the HOT Raw Data API, filtered on the highway tag. They are LineString centrelines with full OSM tags preserved per feature under the tags property. Source projects are validated to at least 95% in the Tasking Manager, so the road network was deliberately mapped and reviewed rather than incidental.

No raster mask is shipped. A centreline has no width, so rasterizing means choosing a buffer, and the right buffer depends on road class: a trunk road and a footway are not comparable. Burning one value would bake that choice in and discard it. Buffer the centrelines yourself, at the widths your task needs:

import geopandas as gpd

WIDTHS = {"trunk": 10.0, "primary": 8.0, "residential": 5.0, "path": 1.5}

roads = gpd.read_file(path)
roads["highway"] = roads["tags"].map(lambda t: t.get("highway"))

metric = roads.to_crs(roads.estimate_utm_crs())
metric["geometry"] = metric.buffer(metric["highway"].map(WIDTHS).fillna(3.0))

Classes

The highway tag value is the road class, carried per feature in label_geojson and summarised per tile in class_counts, following OSM Key:highway. The measured distribution is in the generated stats below.

Classes are what contributors chose, not an enforced schema. None of the source projects set idPresets, so no tag vocabulary was imposed at mapping time. Expect an uneven tail and group classes to suit your task rather than training one head per value.

Composition and known biases

Read these before benchmarking on this dataset.

Major roads are a small minority. The distribution is dominated by residential, which alone is 63.5%, followed by track, service, unclassified, and path. Trunk, primary, secondary, motorway and their link roads together are 3.9%, and most of those come from the two Libyan projects. This is primarily an unpaved and informal road extractor for high-resolution drone imagery; treat the major-road classes as too sparse to train or evaluate on in isolation.

Two Libyan flood-response projects supply about 60% of tiles. They cover large sparse areas, so their road density is well below the rest of the dataset. Filter on country or project_id if you need a geographically balanced sample, and report per-country results rather than a single aggregate.

Empty tiles are subsampled. Tiles containing no road are capped per project against the number of labelled tiles, so the published mix is closer to balanced than the ground truly is. Every labelled tile is retained; only empty ones are sampled, with a stride across the project so they stay spatially spread. The exact counts are in the generated stats. The raw build retains every tile.

Labels and imagery can differ in date. Label completeness follows OpenStreetMap coverage at extraction time, so a road visible in a chip is not guaranteed to appear in the labels.

Splits

Splits are assigned per project, keeping tiles from one project together to avoid spatial leakage. Projects sharing an imagery URL cover the same ground and are assigned as one group.

Sources and licensing

Component	Source	License
Imagery	OpenAerialMap	CC-BY 4.0, attribute contributors of Open Imagery Network
Labels	OpenStreetMap	ODbL 1.0, attribute OpenStreetMap contributors
Project metadata	HOT Tasking Manager

Provenance

Built with tm-oam-ds-builder. Per-project metadata, areas of interest, and aggregate statistics ship alongside the Parquet.

Generated Stats

• Dataset version: 0.1.0
• Tile pairs (num_examples): 61,719
• Total projects: 45
• Feature type: highway
• Tiles with labels: 33,691
• Tiles without labels: 57,042
• Total label features: 66,309
• Avg roads per tile: 0.7
• Total area: ~683.3 sq km
• Countries: 21
• Generated at: 2026-08-06T05:54:43.855915+00:00

Label classes (OSM highway tag)

Class	Features	Share
residential	42,136	63.5%
track	5,730	8.6%
service	4,807	7.2%
unclassified	4,463	6.7%
path	2,601	3.9%
tertiary	2,127	3.2%
footway	1,753	2.6%
trunk	1,068	1.6%
primary	600	0.9%
secondary	510	0.8%
trunk_link	299	0.5%
motorway	51	0.1%
primary_link	46	0.1%
steps	31	0.0%
living_street	28	0.0%
... 7 more

Coverage by Country

Country	Projects	Tiles	Roads	Area (sq km)
Libya	2	65,429	33,197	392.2
Eswatini	3	10,695	11,305	106.8
Mozambique	6	4,992	7,188	39.4
Peru	2	4,623	7,439	85.3
Nepal	2	930	1,624	12.1
Kenya	2	659	1,457	8.3
Liberia	5	537	545	6.5
Tajikistan	2	526	162	4.5
Mexico	2	422	422	3.3
South Africa	1	323	934	2.9
Tanzania	4	285	292	4.2
Colombia	1	193	78	5.5
Iraq	1	188	137	1.3
Ghana	3	182	248	2.8
Uganda	3	173	44	2.1
Trinidad and Tobago	1	134	106	1.8
Madagascar	1	119	157	1.0
Japan	1	113	435	0.8
Philippines	1	97	144	1.0
Argentina	1	62	26	0.5
Sierra Leone	1	51	369	0.8