jp-admin-2026-09

Japan's 47 prefectures and 1,918 municipalities, each with its codes, its name in kanji, kana and Latin script, its polygon, its population and its households. Two rungs and nothing finer.

CC BY 4.0. That is the point. Japan's openly licensed geography has been split between OpenStreetMap, which has polygons and is ODbL, whose share-alike reaches everything built from it, and the government's own registries, which are CC BY but publish either names without boundaries or boundaries without names. This is both, under attribution alone.

from datasets import load_dataset

pref = load_dataset("yuiseki/jp-admin-2026-09", "prefectures", split="train")
muni = load_dataset("yuiseki/jp-admin-2026-09", "municipalities", split="train")

What is here

	rows
prefectures	47
municipalities	1,918	including 171 wards of the 20 designated cities

923 municipalities belong to a 郡. Every row has a representative point.

column
lg_code, code5, pref_code	the registry's six-digit code, its five-digit form and the prefecture
pref, county, city, ward	each with _kana and _roma
name, name_roma	松山市 and Matsuyama-shi; for a ward, 浜松市中央区
efct_date	when the municipality came into being
rep_lon, rep_lat	the registry's representative point
population, households	the 2020 census, summed from small areas
small_areas	how many were summed
geometry_source	census small areas, union of its wards, or none
geometry	WKB polygon, JGD2000 longitude and latitude (EPSG:4612)
bbox	the polygon's bounds, xmin, ymin, xmax, ymax; null where geometry is

The national population sums to 126,146,099, which is the published total of the 2020 census. That is the check worth trusting: if the small areas had been grouped by the wrong code, or a designated city had taken both its own areas and its wards', it would not come out.

A ward and its parent city cover the same ground, so a national sum must be taken over the rows where ward is null. The prefecture table has done that already.

How the files are laid out

Both files are GeoParquet 1.1. The geo metadata names geometry as WKB, gives its types, its overall bbox and its coordinate reference system as PROJJSON, EPSG:4612, JGD2000 longitude and latitude. That is the datum the census declares in the .prj of all 47 of its archives, and nothing here reprojects. GeoPandas, QGIS and DuckDB pick the geometry and the CRS up without being told.

Rows are sorted by pref_code, then lg_code, and each prefecture is one row group, in both files: 47 row groups each. The Parquet statistics of pref_code, code5 and lg_code then say which row group a code is in, and a reader that filters on one fetches that prefecture and nothing else. Fetching 千代田区 reads Tokyo's 2.7 MB rather than the whole 149 MB file; the largest row group is Nagasaki's 11 MB, all those islands.

-- DuckDB, over HTTP: one row group of 47
SELECT name, population, geometry
FROM 'hf://datasets/yuiseki/jp-admin-2026-09/municipalities.parquet'
WHERE lg_code = '131016';

bbox is the GeoParquet covering column for geometry. WKB carries no statistics, so a filter by area works on bbox.xmin and the rest, and skips the prefectures that the area misses. Its values are the polygon's own bounds, which geometry already implies; it adds 0.1 MB.

Where each column comes from

Two sources, each supplying what the other lacks.

yuiseki/abr-src-2026-09, the Digital Agency's Address Base Registry, gives the codes, the names in three scripts, the county and ward structure, the effective dates and the representative points. It has no boundaries.

yuiseki/estat-boundary-2020, the 2020 census small-area boundaries, gives the polygons, the population and the households. Its names are the census's own and its geometry is at a finer rung.

They join on the five-digit municipality code, which is the first five digits of lg_code and the PREF + CITY of the census. 1,889 of 1,918 join directly. The rest are three named cases rather than a residue.

20 designated-city parents. The census has 札幌市中央区 and not 札幌市, so the city is the union of its wards. geometry_source says so.

6 villages of the Northern Territories. 色丹村, 泊村, 留夜別村, 留別村, 紗那村, 蘂取村. The registry lists them; the census does not survey them. They have a name, a code and a representative point, and null for geometry, population and households.

3 wards Hamamatsu created in 2024. 中央区, 浜名区, 天竜区, after the 2020 census. Hamamatsu itself has a polygon, built from the seven wards it had then; its three current wards do not.

Null rather than zero in all nine cases. A village of nobody and a village nobody counted are not the same thing.

Two vintages in one table

The names and codes are the registry as of 2026-09, and the boundaries and counts are the census of 2020. Six years apart, and the three Hamamatsu wards are what that gap looks like. Any municipality that changed between them will show it the same way: the current name over the older shape.

This is stated rather than smoothed. The alternative, dropping to the 2020 municipality list, would mean publishing names that no longer exist.

What it is not

It stops at the municipality. The 町字 are in abr-src-2026-09, all 727,428 of them, and the census small areas are in estat-boundary-2020, all 232,019. Both are published whole, so a third copy of them here would be a liability rather than a convenience.

It is not a cartographic product. The polygons are the census's small-area boundaries dissolved, at the resolution the census publishes, unsimplified and unprojected.

Reproducing it

git clone https://github.com/yuiseki/jp-admin-2026-09
cd jp-admin-2026-09
python3 scripts/01_build.py      # reads the two datasets at pinned revisions
python3 scripts/02_verify.py

Both inputs are pinned by commit sha in scripts/01_build.py, and both carry the publisher's own archives, so the chain can be checked to the byte on either side.

Licence

CC BY 4.0.

出典：「アドレス・ベース・レジストリ」（デジタル庁）
出典：「令和2年国勢調査 小地域（町丁・字等別）境界データ」
（政府統計の総合窓口(e-Stat)）

The Digital Agency applies PDL1.0 and e-Stat applies 政府標準利用規約 （第2.0版）; both state compatibility with CC BY 4.0 and neither is share-alike. LICENSE lists the six modifications in Japanese, as both sets of terms ask.