worldpop-jp-2026-01

WorldPop's gridded population of Japan, 2015 to 2030, release R2025A: how many people live in every 100 m and every 1 km cell, by sex and five-year age band at 1 km, and the degree of urbanisation of every cell. 1,024 rasters as WorldPop published them, and the same rasters as Cloud Optimized GeoTIFF.

CC BY 4.0. Cite each product by its DOI (see Licence).

Why this exists

WorldPop's server sends whole files and ignores HTTP Range requests, at about 400 KB/s a connection, and its GeoTIFFs are striped. To look at one city you fetch the country. The COGs here are tiled at 512 with overviews, and the Hub answers Range requests, so GDAL's /vsicurl/ or rasterio can read a window.

The files of a release are not frozen: within R2025A v1 the population rasters carry dates from July and August 2025 and the degree of urbanisation from January 2026. The month in the name is the latest of them. Every file here has the sha256 recorded for it when it was first fetched, on 2026-09-28.

What is here

raw/GIS/... holds WorldPop's files under WorldPop's own paths; cog/GIS/... the COGs under the same paths. The age bands start at 0 (under 1), 1 (1 to 4), then 5, 10, ... 85, and 90 for 90 and over.

The files config, parquet/files.parquet, lists all 1,024 with their product, year, sex, age band and level, both paths, sizes, sha256, and the count and sum of their valid pixels. The sum of a population raster is the number of people it places, so the table answers questions without opening a raster:

-- DuckDB, straight off the Hub: people of 80 and over, by year
SELECT year, round(sum("sum")) AS people
FROM 'hf://datasets/yuiseki/worldpop-jp-2026-01/parquet/files.parquet'
WHERE project = 'Age and Sex Structures' AND sex = 'total' AND age_from >= 80
GROUP BY year ORDER BY year;
-- 2015  9,807,304   ...   2030  15,422,003

These are WorldPop's estimates and projections, not the census. The sexes and ages of a year do not add up to its total exactly: in 2020 the 60 age and sex rasters place 126,475,186 people and the 1 km total 126,574,853, a difference of 0.08%. They are separate models.

Reading a window of a COG over HTTP:

import rasterio
from rasterio.windows import from_bounds

url = ("https://huggingface.co/datasets/yuiseki/worldpop-jp-2026-01/resolve/main/"
       "cog/GIS/Population/Global_2015_2030/R2025A/2020/JPN/v1/100m/constrained/"
       "jpn_pop_2020_CN_100m_R2025A_v1.tif")
with rasterio.open(url) as r:
    taito = r.read(1, window=from_bounds(139.76, 35.69, 139.81, 35.73, r.transform))

How the COGs were made

DEFLATE with a predictor, 512 by 512 tiles, overviews made automatically. The degree of urbanisation holds class codes, so its overviews use nearest neighbour; averaging would show codes that do not exist. Every value at full resolution was compared with the original and is equal. The overviews are new and are the only thing in a COG that is not in its original.

Reproducing it

git clone https://github.com/yuiseki/worldpop-jp-2026-01
cd worldpop-jp-2026-01
uv run python scripts/01_download.py      # 1,024 files, each against its recorded sha256
uv run python scripts/02_verify.py        # bytes, pixels and people; 16 years of each product
uv run python scripts/03_export_cog.py    # COGs, every value compared
uv run python scripts/04_catalog.py

The download took two hours on 2026-10-02, most of it the sixteen 100 m rasters at 15 to 20 minutes each.

Licence

CC BY 4.0, as WorldPop publishes these datasets. Cite each product as WorldPop gives it; LICENSE beside this file has the four citations in full:

The originals under raw/ are WorldPop's bytes. The COGs under cog/ change the form, not the values, and LICENSE says so.