yhay81/japan-municipal-open-data-atlas-2026

Japan Municipal Open Data Atlas 2026

Release status: public release 2026.08.8. This release was approved after the recorded provenance, reproducibility, and cross-platform checks passed.

Official statistics for every Japanese municipality, already joined, typed, and documented — plus the name-matching table you would otherwise have to build yourself before any of it is usable.

Also distributed as a Kaggle dataset mirror with runnable notebooks, browsable without installation in the Municipality Explorer Space, and grouped in the Atlas collection.

The gap this fills

Japanese municipal statistics are published as separate Excel workbooks per statistic, in Japanese, keyed by an area code that most real-world datasets do not carry. The usual first week of any Japan-wide regional analysis is spent on the same five problems, none of which are about the analysis:

1. Joining by name fails. 中央区 names 11 different areas, 南区 names 12. Within one prefecture, 神奈川県南区 is still ambiguous between Yokohama and Sagamihara. A name join that looks like it worked is the most common silent defect in this domain.
2. Orthography drifts. The official spelling is 龍ケ崎市, 茅ヶ崎市, 飯舘村, 塩竈市; published datasets write 竜ケ崎市, 茅が崎市, 飯館村, 塩釜市. All of them are the same municipality.
3. The official full name is not the everyday name. e-Stat records 北海道石狩振興局当別町; every other dataset says 北海道当別町.
4. Designated cities are counted twice. A designated-city total row and its ward rows are both present, so a naive national sum overstates population. Tokyo has the same trap in the other direction: the special-wards total row 13100 overlaps its 23 ward rows, which are themselves municipalities.
5. Codes expire. A dataset from 2010 carries codes that were merged away years ago, and joining it to a current master silently drops those rows.

This release ships all five resolutions as data, with a deterministic generator and tests, so the joining problem is solved once rather than in every notebook.

Tables

Analysis tables:

• municipalities_city_level: 1,747 non-overlapping city-level rows. This is the safe default for aggregation and comparison; it excludes designated_city_ward rows and the Tokyo special-wards total 13100 (special_ward_area_total) while retaining designated-city totals and the 23 special wards, so a straight column sum matches the official national totals.
• municipalities: 1,919 rows keyed by the official five-digit standard_area_code. It contains the municipality master, census, and resident-register fields.
• municipalities_ward_level: 1,898 rows — the finest non-overlapping partition. It keeps designated-city wards and drops the designated-city totals and the Tokyo special-wards total, so ward-level urban analysis sums to the same official national totals as the city-level view.
• ssds_indicators: 26,866 long-form rows keyed by standard_area_code and indicator_code. Every value carries its actual observation year and bilingual unit.
• ssds_indicators_wide: the same 14 indicators pivoted to 1,919 rows and one column per indicator. Each column name carries the actual observation year, so live_births_2023 cannot be mistaken for a 2026 value. Nothing is imputed: ssds_missing_value_count and ssds_missing_values state exactly which indicators are null and why.

Time-series tables:

• resident_register_annual: 24,947 rows, one per area per survey year, covering 2014–2026. Every year shares one definition — population observed on 1 January, dynamics over the whole preceding calendar year, and the total-resident scope that counts Japanese and foreign residents together — so the years are comparable without further alignment. Earlier years of the same official series are not included: 2013 observes on 31 March over a fiscal year, and 2012 and earlier count Japanese residents only. Published codes are followed forward to the code in force in this release through municipality_code_changes; where a reorganization leaves more than one successor, the row is dropped rather than attributed to a guess.
• resident_register_population_wide: the same total population pivoted to 1,919 rows and one column per survey year, for a quick trend read. 1,915 areas have all 13 years; the rest state their missing years rather than carrying an imputed value.
• resident_register_age: 69,084 rows of population by five-year age class, one row per area per survey year per published sex row, covering 2015–2026. The source publishes the total and the two sexes independently, so sex = total is an official value rather than male plus female. Two columns say where the parts do not reach the whole: age_unaccounted counts residents the twenty-one classes do not cover, because the source withholds age detail for foreign residents in some municipalities and counts residents of unknown age in the total only; age_sex_unspecified counts residents the male and female rows do not cover, which first appears in the 2025 edition. Decide which denominator an age share uses before computing one. 2014 is excluded because it stops at a single 80-and-over class.

Two caveats travel with these tables. area_is_smaller_than_current marks the 1 area-year where a separately published municipality was later merged in, so the row measures less territory than its code covers today. population_is_masked and dynamics_are_masked mark the halves the source itself withholds in a reorganization year: an area retired before the observation date keeps its dynamics but has no population, and an area created on that date has population but no preceding-year dynamics.

Join tables:

• municipality_join_keys: 1,919 rows, one per area, carrying every identifier form in one place — the five-digit standard_area_code, the six-digit local_government_code, the district and designated-city name split apart, the everyday municipality_name_ja, kana, official English, and flags for whether each name is unique nationally and inside its prefecture.
• municipality_code_changes: 1,437 rows of the official municipality-code revision list, covering every change in force between 2005-04-01 and 2024-01-01. current_standard_area_code follows each withdrawn code forward — through more than one merger where necessary — to the code in force in this release, so a 2010 spreadsheet can be joined to today's master. 59 rows stop short of a current code because the source explains those splits only in prose; nothing is guessed.
• municipality_name_lookup: 16,246 rows mapping 15,828 normalized name spellings to area codes. 15,569 keys resolve to exactly one area; the remaining 259 carry match_count and is_unique = false so an ambiguous name fails loudly instead of joining to the wrong municipality. Every one of the 1,919 areas is reachable through at least one unambiguous key.
• municipality_historical_name_lookup: 3,022 rows mapping 2,923 normalized spellings of the 978 municipality names retired by the official code revisions in force between 2005-04-01 and 2024-01-01 to the code in force in this release. A 2010 spreadsheet that still says 妙高高原町 resolves to today's 妙高市 with the retirement date attached; key_is_also_a_current_name flags the 157 spellings that are also in use today so nothing matches silently, and names whose split the source explains only in prose carry has_current_code = false rather than a guess.

All twelve tables are supplied as CSV and Parquet with the same logical values.

Quickstart / クイックスタート

Load any table directly from Hugging Face:

from datasets import load_dataset

city_level = load_dataset(
    "yhay81/japan-municipal-open-data-atlas-2026",
    "municipalities_city_level",
    split="train",
)
ssds_wide = load_dataset(
    "yhay81/japan-municipal-open-data-atlas-2026",
    "ssds_indicators_wide",
    split="train",
)
annual = load_dataset(
    "yhay81/japan-municipal-open-data-atlas-2026",
    "resident_register_annual",
    split="train",
)

On Kaggle, start with the ready-to-use non-overlapping view:

import pandas as pd

root = "/kaggle/input/japan-municipal-open-data-atlas-2026"
city_level = pd.read_parquet(f"{root}/municipalities_city_level.parquet")

Attaching official codes to your own municipality names

Normalize your name column the same way the lookup keys were normalized, then join. is_unique keeps an ambiguous name from silently matching the wrong area.

import unicodedata

import pandas as pd

lookup = pd.read_parquet(f"{root}/municipality_name_lookup.parquet")

# The same rule that produced lookup_key: NFKC, drop whitespace, fold reviewed
# variant characters. FOLD is published in schema.json and reproduction.zip.
FOLD = {"龍": "竜", "舘": "館", "檜": "桧", "竈": "釜", "﨑": "崎", "濱": "浜"}

def normalize(name: str) -> str:
    text = "".join(unicodedata.normalize("NFKC", name).split())
    return "".join(FOLD.get(character, character) for character in text)

mine = pd.DataFrame({"city": ["北海道当別町", "神奈川県茅が崎市", "福島県飯館村"]})
mine["lookup_key"] = mine["city"].map(normalize)

matched = mine.merge(
    lookup.loc[lookup["is_unique"], ["lookup_key", "standard_area_code"]],
    on="lookup_key",
    how="left",
)
unmatched = matched.loc[matched["standard_area_code"].isna(), "city"]

A name left unmatched is either ambiguous (look it up in the full lookup table and disambiguate with the prefecture) or refers to a municipality that no longer exists under that name. The lookup table is a snapshot of the codes in force on the source date; for a name retired by a revision since 2005-04-01, continue with municipality_historical_name_lookup, which maps retired spellings to the code in force today and flags spellings that are also current names.

The public 10-minute Kaggle quickstart demonstrates safe geographic filtering, 2020-2025 population change, natural versus social population change, and an observation-year-aligned SSDS join. The public population-change analysis shows how five-year Census direction and calendar-2024 natural/social change can be described without ranking, causal, or forecast claims.

Questions this release supports

• How do I attach official area codes to a spreadsheet that only has municipality names, and know which rows I could not match safely?
• How many city-level areas grew or declined between the readjusted 2020 Census and preliminary 2025 Census?
• How do natural and social population changes combine across municipalities in the 2024 resident-register period?
• How can one join a selected SSDSE indicator while retaining its observation year, unit, source, and missingness reason?

日本語概要

全国1,919地域の標準地域コードを軸に、2025年国勢調査速報、住民基本台帳、 SSDSEの14指標を結合した分析用データです。値ごとの観測時点、速報・確報の 区分、単位、欠損理由を保持しています。集計には、指定都市行政区と東京都特別 区部計 (13100) の重複を除いた1,747地域の municipalities_city_level を既定として利用できます。単純合計が公式の全国値と 一致します。完全な地域表で市計・区部計と行政区を併用する場合は、分析単位を 明示してください。政令指定都市を行政区単位で扱う場合は、市計と特別区部計を 除いた最細の非重複ビュー municipalities_ward_level (1,898地域、単純合計は市レベルと同じ公式全国値) を利用 できます。2025年国勢調査は速報値であり、公式判断、因果推論、 将来予測には使用できません。

自治体名から標準地域コードを引くには municipality_name_lookup を使います。 NFKC 正規化、空白除去、異体字の畳み込みを施した表記を鍵にしており、龍ケ崎市 と 竜ケ崎市、茅ヶ崎市 と 茅が崎市、飯舘村 と 飯館村、塩竈市 と 塩釜市 は同じ鍵に解決します。異体字を含む公式表記 (龍ケ崎市 など) は 畳み込み前の形でも鍵に含まれるため、公式名そのままの完全一致でも解決します。 中央区 のように全国で重複する名称は is_unique = false と match_count で明示されるため、誤結合が黙って通ることは ありません。2005年4月以降の改正で消滅した旧自治体名 (妙高高原町など) は municipality_historical_name_lookup で現行コードへ解決できます。現行名と 同綴りの旧名には明示フラグが付き、後継が一意に定まらない廃止は推測せずその旨を 記録します。municipality_join_keys には5桁コード、6桁の全国地方公共団体コード、 郡・振興局と指定都市の分離、ふりがな、英語名がまとまっています。

Provenance

The source files are official e-Stat releases from the Statistics Bureau of Japan and the Ministry of Internal Affairs and Communications. sources.json records source URLs, retrieval timestamps, publication timestamps, file sizes, and SHA-256 checksums. This project edited and processed those source tables; e-Stat is not responsible for the transformed output.

Source code revision: bd958ac82f5953fe7698b28653d661ba75ccb612. Schema version: 1.7.0.

Coverage and missing values

• Census observations are absent for six northern-territory reference codes and remain null, not zero.
• Resident-register table 25-03 omits Tokyo special-wards total 13100; its 23 ward rows remain available and no total is synthesized.
• The resident-register source explicitly publishes zero values for the six northern-territory reference rows; these zeroes are preserved.
• SSDS source symbols and normalized missingness reasons are retained. Economic values published only for designated-city totals are not allocated to wards.

designated_city_total rows overlap their designated_city_ward rows, and the special_ward_area_total row 13100 overlaps the 23 Tokyo special wards. Filter statistical_area_role before aggregation to avoid double counting, or start from municipalities_city_level, which already excludes both overlaps.

Derived tables

municipality_join_keys, municipality_name_lookup, and ssds_indicators_wide add no new official values. Every field is copied from a validated release row or computed by a rule that the builder verifies against values already in the release:

• The six-digit local_government_code is computed with the official JIS X 0402 check-digit rule and compared against all 1,918 codes published in the resident-register source; the build fails on any disagreement. 1 code is supplied by the rule alone, for the one area the source omits, and local_government_code_source records that.
• Lookup keys apply NFKC normalization, whitespace removal, and a reviewed variant-character map, then expand interchangeable connective characters (ヶ ケ ヵ カ ガ が 箇 and 之 ノ の). Official spellings that contain a variant character — such as 龍ケ崎市, 飯舘村, and 塩竈市 — are kept as keys unfolded as well, so exact official names resolve without any client-side normalization. The map deliberately excludes 嶋 → 島, which would merge 茨城県鹿嶋市 with 佐賀県鹿島市. A test asserts that folding merges no two areas that were not already ambiguous under their plain names.
• ssds_indicators_wide is a pure pivot: a test compares every cell against the long table.

Intended use

Suitable for reproducible descriptive analysis, education, public-data quality work, and transparent non-high-stakes regional comparisons. Values with different observation dates or statistical definitions must not be treated as interchangeable.

Limitations

The 2025 census values are preliminary. This dataset does not provide official rankings, causal conclusions, forecasts, eligibility decisions, legal-address evidence, or support for high-stakes individual decisions.

Quality and reproducibility

The release builder validates raw-source checksums, exact code-based joins, primary keys, expected coverage, and reviewed source-specific quality totals. CSV and Parquet exports must be logically equal under the published Arrow schema. quality.json, schema.json, sources.json, manifest.json, and checksums.sha256 provide machine-readable audit evidence.

reproduction.zip.bin contains the source code, fixed Python environment, tests, configuration, audit notebook, and step-by-step instructions used for this candidate. Official raw files are reacquired from the URLs recorded in the source metadata and are not redistributed in the archive. The ZIP bytes use a .bin suffix in platform bundles to prevent automatic extraction; rename the downloaded file to reproduction.zip if desired.

Maintenance

The dataset is reviewed at least annually after the relevant official releases. An earlier update may be issued for a material source correction or schema defect. Preliminary census values will be replaced only by an explicitly versioned release after final values are published and the same provenance and quality checks pass. Previously published versions remain documented; a release is retired rather than silently changed if its source is withdrawn or cannot be verified reproducibly.

License and attribution

Source numerical tables are reused under the e-Stat Terms of Use, which are compatible with CC BY 4.0. Attribute e-Stat and the producing ministry/statistic, and state that the source was edited or processed when redistributing derivatives.

The code in reproduction.zip.bin is provided under the MIT License.

Suggested citation

Hayashi, Yusuke (2026). Japan Municipal Open Data Atlas 2026, release 2026.08.8, schema 1.7.0. Kaggle and Hugging Face. Also cite e-Stat and the producing ministry/statistic for reused source values.

Development disclosure

AI assistance was used in code and documentation development. Release claims are backed by deterministic tests and recorded official-source evidence; this does not imply independent verification of the underlying official statistics.