NoeFlandre/georeset-text-label-benchmark

GeoReset text-label benchmark: Description/EUNIS overlap

Author: Noé Flandre

This dataset pairs exact Description sentence occurrences whose land-use relevance label is yes with an existing EUNIS assignment on the corresponding OpenStreetMap polygon. It is a reproducible overlap and coverage artifact. It does not evaluate predictions, train a model, fill missing labels, or recompute geometry.

Pinned inputs

• Description relevance labels: NoeFlandre/osm-polygon-description-tag-landuse, revision d52781febf61a1ee77e4aceb02d9b607628b37c5.
• Description values and EUNIS polygon assignments: NoeFlandre/osm-polygon-description-tag-eunis, revision 88eabd83fcc60cb67ab4c1768a16e3d05e25ffb3.
• Shared input revision: b4706eb315c66a8a57135289e7612dca8e03caf8.
• EUNIS reference: EEA EUNIS habitat probability maps v1 (2021). The retained polygon rows' source-version fields were checked against this pinned reference.

The full provenance, EEA factsheets, license notes, computation commit, validation commit, and artifact SHA-256 digests are in manifest.json and the project provenance documentation.

Join contract

Labels join to Description records on (description_identity, tag_key). The zero-based sentence_index selects an exact sentence occurrence. Description records join to polygon rows on (osm_type, osm_id) and source_pbf must also match. Sentence SHA-256 values verify text integrity and are not join keys; identical text at different places remains distinct.

Coverage

All 386 aligned shards were scanned. Baseline counts matched the pinned source summaries. EUNIS assigned means the corresponding polygon already had an EUNIS label in the pinned input. Missing assignments were retained in the coverage report but are not included in overlap.parquet.

Label outcome	EUNIS assigned	EUNIS missing	Total
yes sentence rows	224,789	500,582	725,371
no sentence rows	94,843	146,632	241,475
failed rows	4,716	7,211	11,927
skipped_unsplit rows	15,166	63,063	78,229

Source stages: 906,631 polygon rows, of which 276,604 have EUNIS assignments; 919,126 Description tag rows; 1,057,002 sentence-label rows; and 978,773 sentence rows. There were 0 polygons without a corresponding Description record in the aligned snapshots.

The retained dataset has 224,789 rows, one per yes sentence occurrence with an EUNIS-assigned polygon. Repeated sentences are counted separately by occurrence. summary.json provides further counts by EUNIS code, language, and tag key. overlap.parquet is Zstandard-compressed Parquet.

Output fields

The Parquet rows contain OSM source and feature keys, Description identity and tag, sentence index and text, text SHA-256, language code, and the existing EUNIS code, name, overlap percentage, and source version. See manifest.json for the complete typed schema.

Data terms and limits

OpenStreetMap-derived source data are under the ODbL 1.0; attribute OpenStreetMap contributors and preserve applicable share-alike obligations. The EEA EUNIS probability-map collections identify CC-BY 4.0; attribute the European Environment Agency and retain the source links in the manifest and provenance documentation. Project code is Apache-2.0; that code license does not replace the source-data terms.

EUNIS is polygon context, not sentence-level ground truth. Scientific validation of the final EUNIS assignments is unconfirmed, and this dataset makes no claim that an assigned habitat is a valid label for each individual sentence.