wiki-geoparquet

English Wikipedia main-namespace articles with Earth coordinates, as GeoParquet + PMTiles.

Coordinates from Wikidata P625, enriched with instance type, country, population, GeoNames IDs, inlink counts, and more. Updated monthly from Wikidata + Wikipedia dumps.

demo: shane98c.github.io/wiki-geoparquet

Data

All files are on R2 with CORS enabled — query directly from the browser or any Parquet-aware tool:

File	Description
wikipedia_geotagged.parquet	GeoParquet, Hilbert-sorted with bbox covering
wikipedia_geotagged.pmtiles	Vector tiles, auto-zoom with overzoom, drops by article length
wikipedia_search.parquet	Lightweight search index (lowercased label, coords, inlink count), sorted by label for prefix-range row-group pruning

Pinned versions are also available on GitHub Releases.

Schema

Column	Type	Description
geometry	WKB Point	WGS84 coordinates (from Wikidata P625)
page_id	int32	Wikipedia page ID
qid	string	Wikidata QID (e.g. Q90)
label	string	Article title
description	string	Short description from wikibase-shortdesc
instance_of	string	Wikidata P31 type label (e.g. "city", "mountain", "museum")
instance_of_qid	string	Wikidata P31 QID (e.g. "Q515", "Q8502") — for downstream mapping to custom taxonomies
country	string	Country name from Wikidata P17
population	int64	Population from Wikidata P1082 (nullable)
geonames_id	string	GeoNames ID for cross-referencing
elevation	int16	Ground elevation in meters, sampled from Copernicus DEM
gt_type	string	Wikipedia geo classification (supplementary, from geo_tags)
page_len	int32	Article length in bytes
inlink_count	int32	Number of namespace-0 pagelinks pointing here
wikipedia_url	string	Full article URL
image_url	string	Wikimedia Commons image URL
related_images	list<string>	Additional Wikidata images (P6802)
bbox	struct	Covering bbox for spatial predicate pushdown

The PMTiles carry page_id, label, inlink_count, and instance_of_qid — the QID lets downstream consumers apply their own category taxonomy via Mapbox/MapLibre match expressions without rebuilding tiles. The demo enriches popups on click by querying the GeoParquet directly via DuckDB-WASM. Reconstruct article URLs client-side: https://en.wikipedia.org/?curid={page_id}

Quick start

Query with DuckDB

INSTALL spatial; LOAD spatial;

-- Find the most notable geotagged articles
SELECT label, instance_of, country, inlink_count, ST_AsText(geometry)
FROM 'https://pub-016504dd3a4d419a9c17a8939840935e.r2.dev/v2/wikipedia_geotagged.parquet'
ORDER BY inlink_count DESC
LIMIT 20;

-- Articles near Paris (48.8566, 2.3522) — ~50km bbox
SELECT label, instance_of, country, inlink_count
FROM 'https://pub-016504dd3a4d419a9c17a8939840935e.r2.dev/v2/wikipedia_geotagged.parquet'
WHERE bbox.xmin BETWEEN 1.7 AND 3.0
  AND bbox.ymin BETWEEN 48.4 AND 49.3
ORDER BY inlink_count DESC
LIMIT 20;

Browser search with duckdb-wasm

The wikipedia_search.parquet index has columns label (lowercased, used for both sorting and lookup), lon, lat, and inlink_count for ranking. It's sorted by label so a lowercased prefix range lets DuckDB skip row groups and fetch ~2 MB per query instead of the full ~25 MB file.

// In duckdb-wasm, INSTALL httpfs (or SET builtin_httpfs = false) — without
// this the built-in HTTP handler downloads the whole file on every query.
// See https://github.com/duckdb/duckdb-wasm/issues/2153.
await conn.query("SET builtin_httpfs = false;");

const q = userInput.toLowerCase().replace(/'/g, "''");
await conn.query(`
  SELECT label, lon, lat
  FROM 'https://.../wikipedia_search.parquet'
  WHERE label >= '${q}' AND label < '${q}~'
  ORDER BY inlink_count DESC
  LIMIT 10
`);

Use pmtiles in MapLibre

import { Protocol } from "pmtiles";

let protocol = new Protocol();
maplibregl.addProtocol("pmtiles", protocol.tile);

const map = new maplibregl.Map({
  style: {
    sources: {
      wikipedia: {
        type: "vector",
        url: "pmtiles://https://pub-016504dd3a4d419a9c17a8939840935e.r2.dev/v2/wikipedia_geotagged.pmtiles",
      },
    },
    layers: [
      {
        id: "articles",
        source: "wikipedia",
        "source-layer": "wikipedia",
        type: "circle",
        paint: {
          "circle-radius": [
            "interpolate",
            ["linear"],
            ["sqrt", ["get", "inlink_count"]],
            0,
            2.5,
            280,
            20,
          ],
          "circle-color": "#4264fb",
        },
      },
    ],
  },
});

How it works

1. Streams the Wikidata truthy N-Triples dump (~70 GB) to extract P625 coordinates and enrichment properties (instance type, country, population, GeoNames ID, images)
2. Resolves P31/P17 QIDs to human-readable labels via the Wikidata API
3. Streams 5 Wikipedia SQL dump files (~12 GB) for article metadata, images, descriptions, and inlink counts
4. Writes Hilbert-sorted GeoParquet with bbox covering via DuckDB spatial
5. Pipes DuckDB to tippecanoe for PMTiles with attribute-based feature dropping

See the Makefile for build steps.

Data source

Coordinates and enrichment from Wikidata entity dumps, article metadata from English Wikipedia SQL dumps. Both released under CC BY-SA 4.0.

License

Code: MIT Data outputs: CC BY-SA 4.0 (derived from Wikipedia/Wikidata)