osm-polygon-selection dataset

A curated set of OpenStreetMap polygons across 104 countries (49 European + 54 in Africa), classified by size bin (small / medium / large, area in [0.1, 100] km²) and tagged by continent (Natural Earth admin0 lookup).

Size bins:

• small — area in [0.1, 1) km² (10,000 m² to 1 km², roughly 100 m × 100 m to ~1 km × 1 km). Examples: a city block, a small park, a single farm field, a small wood lot, a residential courtyard, a parking lot, an industrial yard.
• medium — area in [1, 10) km² (1 km² to 10 km², roughly 1 km × 1 km to 3 km × 3 km). Examples: a large park, a small village/town footprint, a reservoir, a forest patch, an industrial zone, a golf course, a cemetery, a nature reserve.
• large — area in [10, 100] km² (10 km² to 100 km², roughly 3 km × 3 km to 10 km × 10 km). Examples: a large forest, a big lake, an entire town or small city, a large military training area, a national park section, a sizable agricultural region.

Polygons smaller than 0.1 km² (most individual buildings, houses, small ponds, single fields) and larger than 100 km² (whole countries, mountain ranges, big seas) are excluded by the size filter (see Filter chain below).

Status: 100 of 104 countries are clean. 4 country(ies) were killed mid-pipeline (see per_country/<country>/README.md).

Total polygons: 8,010,269 (combined parquet: all_world.parquet).

Layout

This dataset is split across four subfolders so you can pull only what you need:

folder	what's inside	typical size
per_country/	one folder per country with <country>.parquet + README.md	~7 GB total, <1 MB per small country
combined/	all_world.parquet — every polygon in one file	~9 GB
sample/	sample_map.jsonl — ~4k representative polygons for quick viz	<1 MB
preview/	map_preview.png — static map thumbnail	~1 MB

Start with sample/ or preview/ for a quick look. Pull per_country/<country>/<country>.parquet for a single-country study. Use combined/all_world.parquet for cross-country work.

What's in this dataset

Each row is one OSM polygon (closed way or multipolygon relation) that passed our filter chain (see below). The polygon geometry itself is included in the row as WKT (or WKB if OSM_POLYGON_GEOMETRY=wkb is set when the dataset is built) so you can render, query, or reproject it directly without re-deriving from centroid+area.

column	type	description
osm_id	int64	OSM object id (int64).
osm_type	string	OSM object type, "way" or "relation" (string).
centroid_lon	float64	polygon centroid longitude (WGS84, float64).
centroid_lat	float64	polygon centroid latitude (WGS84, float64).
area_km2	float64	polygon area in km² (Web Mercator, float64).
tags	list(string)	OSM key=value tags (list of strings).
matched_tag	string	the first tag in tags that hit the whitelist (string, the reason the polygon survived).
continent	string	Natural Earth admin0 lookup of the centroid (string).
size_bin	string	"small" (0.1-1), "medium" (1-10), or "large" (10-100) km² (string).
country	string	ISO-style country name (string).
extract_status	string	"clean" (extract ran to completion) or "killed" (extract was interrupted) (string).
pbf_date	string	date of the source PBF file (string, from mtime).
geometry_wkt	string	polygon geometry as WKT (WGS84, string). Parse with shapely.wkt.loads(row.geometry_wkt).

Provenance

• Pipeline version: v0.1.0
• Git SHA: 6d90fdf368bb93127478f0517c69344600353066
• Built: 2026-07-03T22:01:20.352249
• Source: Geofabrik regional extracts (https://download.geofabrik.de/)
• Whitelist: 22,075 OSM key=value tags from osm-stats (see docs/whitelist_decisions.md in the project repo, or read the full rationale in the blog post). The whitelist is designed to filter polygons by landuse-style tags (natural, landuse, leisure, amenity, etc.) so the dataset focuses on physical land-cover / land-use features rather than buildings, addresses, or points of interest.

Geographic distribution

(Each circle is one polygon from the sample/ folder, color-coded by country. Circle size is proportional to sqrt(area_km2).)

Size-bin distribution (full dataset)

Counts every polygon in the 8,010,269-polygon dataset by size_bin, computed directly from combined/all_world.parquet via pyarrow.compute.value_counts. Percentages are exact ratios over the entire dataset, not a sample.

size_bin	count	pct
small	6,385,831	79.7%
medium	1,419,867	17.7%
large	204,571	2.6%
Total	8,010,269	100.0%

Example row

Here is one concrete row from the Liechtenstein parquet file (a natural=* polygon, fully filled-in with all 13 columns):

column	value
osm_id	2316236858
osm_type	way
centroid_lon	9.511089
centroid_lat	47.228654
area_km2	0.1143
tags	natural=water
matched_tag	natural=water
continent	Europe
size_bin	small
country	liechtenstein
extract_status	clean
pbf_date	2026-06-26
geometry_wkt	MULTIPOLYGON (((9.5016963 47.2116538, 9.5017011 47.211496, 9.5017372 47.2113713, 9.5017277 47.211...

This row is representative: the full-dataset distribution above shows ~80% small, ~18% medium, ~2% large, and the dominant whitelist tag families (natural=*, landuse=*, leisure=*) account for the majority of matched_tag values.

Filter chain

Each polygon in this dataset has passed three filters:

1. Size filter (Stage 0): area in [0.1, 100] km². Polygons smaller than 0.1 km² or larger than 100 km² are dropped.
2. Whitelist filter (Stage 2): at least one OSM tag in the 22,075-tag whitelist. The whitelist is derived from a clustering of OSM tags across both tfidf and embeddings analyses.
3. Classify (Stage 3): continent assigned via Natural Earth admin0 shapefile, size_bin assigned by area.

Train / val / test split

Every row in every parquet (per_country/<country>/<country>.parquet and combined/all_world.parquet) carries a split column with one of three values: train, val, or test.

split	ratio	polygons
train	80%	0
val	10%	0
test	10%	0
Total	100%	8,010,269

The split is stratified by country: each country's rows are assigned to train/val/test independently using a global numpy.random.default_rng seeded once with 42 and offset per country. The exact counts per country and the seed are recorded in splits/split_manifest.json.

To load only one split (e.g. for training), filter in pyarrow:

import pyarrow.compute as pc
import pyarrow.parquet as pq
table = pq.read_table("combined/all_world.parquet")
train = table.filter(pc.equal(table["split"], "train"))

The split is deterministic and re-runnable:

uv run python scripts/make_split.py            # default seed=42, 80/10/10
uv run python scripts/make_split.py --seed 7   # different reproducible split

Per-country summary

Country	Polygons	Status
albania	14,738	clean
algeria	32,601	clean
andorra	776	clean
angola	19,197	clean
austria	133,711	clean
azores	2,640	clean
belarus	223,750	clean
belgium	125,108	clean
benin	4,614	clean
bosnia-herzegovina	49,715	clean
botswana	5,327	clean
bulgaria	74,567	clean
burkina-faso	8,835	clean
burundi	4,659	clean
cameroon	71,789	clean
canary-islands	2,560	clean
cape-verde	2,417	clean
central-african-republic	53,491	clean
chad	23,667	clean
comores	395	clean
congo-brazzaville	6,643	clean
congo-democratic-republic	0	killed
croatia	47,140	clean
cyprus	4,846	clean
czech-republic	271,062	clean
denmark	175,795	clean
djibouti	576	clean
egypt	24,623	clean
equatorial-guinea	1,004	clean
eritrea	3,278	clean
estonia	47,160	clean
ethiopia	29,663	clean
faroe-islands	1,278	clean
finland	427,870	clean
france	492,538	clean
gabon	3,843	clean
georgia	21,198	clean
germany	1,131,888	clean
ghana	11,445	clean
greece	45,142	clean
guernsey-jersey	670	clean
guinea	12,311	clean
guinea-bissau	2,109	clean
hungary	77,569	clean
iceland	47,896	clean
ireland-and-northern-ireland	159,879	clean
isle-of-man	2,648	clean
italy	276,991	clean
ivory-coast	14,273	clean
kenya	16,916	clean
kosovo	5,377	clean
latvia	47,133	clean
lesotho	19,246	clean
liberia	2,342	clean
libya	13,046	clean
liechtenstein	565	clean
lithuania	76,550	clean
luxembourg	11,460	clean
macedonia	9,330	clean
malawi	5,337	clean
mali	40,392	clean
malta	620	clean
mauritania	9,040	clean
mauritius	1,863	clean
mayotte	606	clean
moldova	35,690	clean
monaco	2	clean
montenegro	11,785	clean
morocco	42,623	clean
mozambique	11,101	clean
namibia	9,316	clean
netherlands	207,459	clean
niger	14,606	clean
nigeria	0	killed
norway	413,801	clean
poland	637,908	clean
portugal	66,287	clean
romania	115,401	clean
rwanda	3,976	clean
saint-helena-ascension-and-tristan-da-cunha	174	clean
sao-tome-and-principe	190	clean
senegal-and-gambia	20,479	clean
serbia	47,189	clean
seychelles	264	clean
sierra-leone	3,366	clean
slovakia	54,888	clean
slovenia	41,526	clean
somalia	48,453	clean
south-africa	0	killed
south-sudan	17,091	clean
spain	240,230	clean
sudan	38,043	clean
swaziland	4,113	clean
sweden	397,661	clean
switzerland	61,156	clean
tanzania	0	killed
togo	3,408	clean
tunisia	8,498	clean
turkey	113,609	clean
uganda	10,449	clean
ukraine	645,578	clean
united-kingdom	205,002	clean
zambia	15,684	clean
zimbabwe	7,545	clean
Total	8,010,269

Back to the dataset root