osm-polygon-description-tag-worldcover

A supervised text to land-cover dataset. Each example pairs OpenStreetMap description and localized-description tag text with the ESA WorldCover class that covers at least 80% of the OpenStreetMap polygon the text describes.

76,601 examples, 74,759 distinct polygons, 76,601 distinct documents.

from datasets import load_dataset

ds = load_dataset("NoeFlandre/osm-polygon-description-tag-worldcover")
print(ds["train"][0]["text"][:200], ds["train"][0]["worldcover_label"])

Geographic coverage map

The map shows the centroids of all 74,759 distinct polygons represented in this release and carrying an ESA WorldCover label. Each point is colored by its WorldCover class. Points are centroids, not polygon boundaries; use lat, lon, and centroid_wkt for the tabular locations.

Example labelled polygons

polygon name	ESA WorldCover class
"a budi"	Tree cover
A Pena Golpilleira	Shrubland
"Erie Express" Train Station	Grassland
"INVEST ZONE" MChJ | Mettalurgik quvur zavodi	Cropland
!PS - Innovationspark Schaumburg	Built-up
"Royal Palace"	Bare / sparse vegetation
Bivacco Brigata Orobica	Snow and ice
't Hilgelo	Permanent water bodies
Beribadh Sonir Hawor Rokkha Badh	Herbaceous wetland
Bocilla Preserve	Mangroves
Agnetehytten	Moss and lichen

One deterministic named polygon is shown for each ESA WorldCover class represented in this release.

What the label means

WorldCover classifies the ground in 10 m pixels. The label therefore describes the land cover of the area containing a feature, not the feature itself: a church is Built-up because its surroundings are, not because the building was classified. Treat this as place-context classification.

Each row carries polygon_area_m2 and observed_fraction so you can restrict to polygons large enough for the dominance test to have been a real filter (polygon_area_m2 >= 2500 is 25+ pixels).

Splits

split	examples	polygons	documents
train	59,986	58,699	59,986
validation	7,658	7,512	7,658
test	8,957	8,548	8,957

Classes

code	class	examples	share
10	Tree cover	19,225	25.1%
20	Shrubland	632	0.8%
30	Grassland	9,108	11.9%
40	Cropland	1,234	1.6%
50	Built-up	43,891	57.3%
60	Bare / sparse vegetation	973	1.3%
70	Snow and ice	31	0.0%
80	Permanent water bodies	1,397	1.8%
90	Herbaceous wetland	42	0.1%
95	Mangroves	20	0.0%
100	Moss and lichen	48	0.1%

Languages (top 15)

language	examples	share
(unspecified)	72,874	95.1%
en	1,306	1.7%
de	596	0.8%
fr	432	0.6%
ru	196	0.3%
it	180	0.2%
es	158	0.2%
nl	91	0.1%
FR	63	0.1%
hu	55	0.1%
ro	54	0.1%
ko	53	0.1%
pl	44	0.1%
pt	29	0.0%
ar	27	0.0%

Geographic coverage

• H3 cells (resolution 5): 19,170
• Source regions: 361
• Bounding box: -179.10867579719636, -54.93691708735652 to 179.5240562102925, 82.52576211486381

Guarantees

Every build is checked against these and fails if any breaks:

• No polygon leakage — no polygon appears in more than one split.
• No document leakage — no document appears in more than one split.
• Geographic blocking — splits are assigned per H3 cell, so nearby places cannot straddle train, validation and test.
• Valid labels — every label is one of the 11 real WorldCover classes; no-data is never a label.
• Dominance — every row's class covers at least 80% of its polygon.
• No exact duplicates — no two rows share both text and label.
• Deterministic — rebuilding the same inputs yields byte-identical files.

Columns

column	meaning
text	Full source text
worldcover_code / worldcover_label	The target class
dominant_fraction	Share of the polygon covered by that class
observed_fraction	Share of the polygon observed at all
polygon_id, osm_type, osm_id	The OpenStreetMap object
lat, lon, centroid_wkt	Polygon centroid
polygon_area_m2	Polygon area
language, project, title, url	The article
split, h3_cell	Split and the cell it was decided on

Provenance

• Source: NoeFlandre/osm-polygon-description-tag at revision fec858b679f5ee7e87f0ecfaaa6b7223b2a7f5e2
• Land cover: ESA WorldCover 2021 v200 (10 m)
• Dominance threshold: 0.8
• Minimum article length: 10 words
• Maximum polygon area: 10000000000.0 m2
• Split seed: 20180101, H3 resolution 5

Code: github.com/NoeFlandre/osm-worldcover

Polygons refused

reason	polygons
below_threshold	210,549
no_valid_class	77
too_large	197

Rows removed after labelling

reason	rows
documents_split_across_splits	0
duplicate_examples	16,708
duplicate_objects_across_regions	0

Licence

Open Database License (ODbL). OpenStreetMap geometry is ODbL. ESA WorldCover is CC BY 4.0.