How data-geo works: WorldPop

WorldPop publishes its population grids as one GeoTIFF per country and year, from servers that ignore range requests: one city means downloading the country (160 MB for Ethiopia, per year). data-geo re-publishes WorldPop R2025A (100 m, 2015–2030, total population and age and sex) as cloud-native files anyone can read by range, with an API that counts people in any place or shape. Built in the open; numbers on this page are measurements, dated.

Data © WorldPop, R2025A v1 constrained, CC BY 4.0. Cite WorldPop for every figure: see licence and citation. data-geo is run by Baobab Tech and is not affiliated with or endorsed by WorldPop.

WorldPop's API

Both serve WorldPop R2025A at 100 m for 2015–2030 under CC BY 4.0, and both count a cell when its centre is inside the area, so their polygon totals agree to under one person. WorldPop launched API v3 (3.0.0) in early November 2026, with a guide for agents and an MCP server at /v3/mcp. Measured 5 November 2026; server times are what each API reports for itself (processing_time_ms, milliseconds).

WorldPop API v3 data-geo
Access Free key, by registration and approval; capped without one No key
Quota 1,000 requests a day anonymous, 10,000 with a key; 60 a minute, 3 at once 2,000 API calls an hour per IP, polygons limited more tightly; no daily quota
Largest polygon 100,000 km² anonymous; 500,000 with a key (up to 1,000,000) Any size up to a whole country (Ethiopia, 1.1 million km², in 3–5 s)
City polygon, server time 42 ms 59 ms median from the cache, 0.45 s on a first read
Same box, 4 years 58 ms 0.1–0.3 s from the cache, 1 s on a first read
Catchment, 3 radii 41 ms about 0.2 s
4° × 4° box (197,000 km²) 413 without a key 24.8 million people, 0.2 s
Age and sex years All 16 2015, 2020, 2025, 2030
Administrative units WorldPop's own (Kenya: 8 former provinces) geoBoundaries ADM0–ADM2, with each boundary's source and licence (Kenya: 47 counties and their sub-counties)
Files GeoTIFFs and zips on data.worldpop.org, no range reads; /raster cuts a GeoTIFF of an area COGs, VRTs, Zarr cube, sum pyramids, Parquet admin totals and STAC, all read by range
How an answer was built Processing time method: stored sums used, tiles read, files, cache hits

WorldPop is faster on small queries (about 40–60 ms against 60–450 ms here); data-geo answers large areas without a key.

On files: WorldPop's API links each country's GeoTIFFs on data.worldpop.org (/countries/{iso3}/files), which answer a range with the whole file, and its /raster cuts a GeoTIFF of an area on request. In October 2026 WorldPop said it plans a GeoZarr release at 1 km and to host its next release on AWS Open Data; on 5 November 2026 the registry listed no WorldPop dataset. When WorldPop's own files can be read by range, this page will say so and point to them.

The same requests to both, 5 November 2026:

Query WorldPop v3 data-geo
Cell at 36.8219, −1.2921, 2025 67.84 67.84
Box over central Addis (38.70–38.80, 8.98–9.06), 2025 1,606,480.21 1,606,479.47
The same box, 2015, 2020, 2025, 2030 in one request 1,266,918.96; 1,435,807.10; 1,606,480.21; 1,780,608.16 1,266,919.47; 1,435,807.72; 1,606,479.47; 1,780,608.09
Change 2015–2030 on that box +513,689.19, +40.55%, 2.3% a year, peak 2030 +513,688.61, +40.55%, 2.30% a year, peak 2030
Indicators, that box, 2025 median age 25.3, sex ratio 93.2, dependency 33.0 (young 28.5, old 4.5), ageing index 15.8, child-woman ratio 22.6, under 5 0.0773, 65+ 0.0339 25.3, 93.18, 33.04 (28.54, 4.51), 15.8, 22.62, 0.0773, 0.0339
Within 1, 5, 10 km of 36.8219, −1.2921, 2025 27,220.88; 937,123.38; 3,205,739.24 27,013.73; 935,401.62; 3,204,498.01
That box, age and sex, 2023 1,537,578.24 in 40 groups 404: age and sex for 2015, 2020, 2025, 2030 only
4° × 4° box around Nairobi (197,000 km²), 2025 413 without a key (limit 100,000 km²) 24,813,185
Kenya's units 8 (the former provinces) geoBoundaries: 47 counties (ADM1) and ADM2 below

Polygon totals agree to under one person in every year: both count a cell when its centre is inside. Catchments differ by 0.8% at 1 km, 0.2% at 5 km and 0.04% at 10 km. WorldPop does not say how it draws a circle; here it is a 256-sided polygon counted by cell centres, and at 1 km the circle's edge crosses about 60 of the roughly 300 cells inside it, so any difference in how the edge is drawn weighs most there.

What you can use

The API at https://data-geo.baobabtech.app/v1/worldpop/:

Call Answers
POST polygon?year= with GeoJSON People inside any polygon, exactly
GET catchment?lon=&lat=&radii=1,5,10&year= People within each distance of a point, and between them
GET point?lon=&lat=&year= People in the 100 m cell under a point
GET totals?iso3=&level=&year= Every administrative unit (geoBoundaries ADM0–2) of a country
GET total?iso3=&unit=&year= One unit, with its boundary's source and licence

Each takes:

Parameter Gives
years=2015-2030 or years=2015,2020,2025 in place of year Every year asked, the change from first to last (absolute, percent, compound annual growth) and the peak year
groups=all, summary or a list (under5, women_15_49, f_20 …) People by sex and age group
indicators=true Median age, sex ratio, dependency ratios, ageing index, child-woman ratio and age shares

Groups and indicators cover 2015, 2020, 2025 and 2030; a request for any other year answers 404 with missing_years.

The files, under https://data-geo.baobabtech.app/worldpop/r2025a/, read by range from GDAL, rasterio, xarray, DuckDB or QGIS. Paths never change; a new release gets a new path. Country codes in file names are lower case (eth.tif).

Path What
100m/{year}/{iso3}.tif Population COG per country and year: float32, EPSG:4326, nodata −99999, 512 × 512 tiles, DEFLATE + floating-point predictor
100m/{year}/global.vrt Every country of a year as one raster
100m.zarr/ Zarr v3 population(time, y, x), all 16 years in each 512 × 512 chunk: a time series in one read
pyramid/{year}/1km.tif, 10km.tif Exact sums of the 100 m grid (sum, count of valued cells; float64)
admin_totals.parquet One row per unit, one column per year, one row group per country
agesex/100m/{year}/{iso3}.tif 43 bands per cell: population, f_00 … f_90, m_00 … m_90, f_under18, m_under18; 32 × 32 tiles, no predictor, empty tiles unwritten
agesex/100m/{year}/global.vrt, agesex/pyramid/{year}/, agesex/admin_totals/{year}.parquet The same three for sex and age
manifest.json Every file's size and SHA-256; per COG, its WorldPop source URL and SHA-256, sum and cell count
stac/catalog.json STAC 1.0: one item per country and year (items/{year}/{iso3}.json), one per year for the VRT and pyramid (items/{year}/global.json)

At the root, /catalog.json (also GET /v1/catalog) says per year whether it is complete, which countries are in and which are confirmed absent; /worldpop/latest.json names the current release. The manifest, STAC and catalogue are generated from the bucket and rewritten as files land; the files they list never change.

Agents: /llms.txt. Tools: /openapi.json. This page as markdown.

Live test

Every endpoint, tested live from Helsinki on 24 October 2026 (the age and sex rows on 1 November), worldwide data.

Working

Test Result
Point, Lagos 2026 237 people in the cell, from Nigeria's file · 0.65 s
Point, Mumbai 2026 152 people · 0.9 s
Polygon, 5 km circle around Lagos 2026 1,169,347 people · 0.7–1.0 s
Age-sex at a point, Nairobi 2025 68 people: 32 female, 36 male, 6 under 5, 19 under 18 · 0.4–0.6 s
Admin totals with age and sex, Kenya's 47 counties 2025 0.6 s
Range read of Nigeria's 2026 file 16 KB in 0.4 s
/, /llms.txt, /openapi.json, how-it-works page and markdown, diagrams all serve

Correctly refused

Request Answer
Point in the mid-Atlantic 404, no country file covers it
Age-sex polygon, worldwide 2026 404, worldwide age-sex not published yet
Polygon for 2024 404, not published yet (downloading now)
Admin totals, default release 404, not built worldwide yet
GeoJSON Point sent as a polygon 400, with the reason

Also live, 2026: central Jakarta, 8,615,605 people in 0.9 s; a box across the Benin–Nigeria border, 30,500,961 in 1.2 s; all of Ethiopia by its boundary, 137,022,972 in 8.2 s. The 2026 sum grids total 8,264,687,130 people, equal to the 242 country files; 13 territories have no WorldPop 100 m file.

How the answers are computed

One counting rule. A 100 m cell belongs to an area when its centre is inside it (even-odd, so holes and multipolygons work). Polygons, admin totals and every age and sex group use it, so a polygon of an administrative boundary returns that unit's total and a breakdown adds up to its total.

One grid, never resampled. WorldPop stores the cell size as 0.00083333333, not exactly 1/1200°. On the stored value every country file starts on a whole cell from (−180, 90); on 1/1200 every one looks slightly off. Every product here is indexed on the stored value, so countries are placed by integer offsets and nothing is resampled (resampling a count grid changes its totals).

Polygons, coarse to fine. Counting a country cell by cell is ~250 million tests. People add up, so a 10 km cell stores exactly the total of the 10,000 cells under it, and a polygon is answered from the sum pyramid:

A polygon over a grid: coarse cells wholly inside it count as one stored sum each, medium cells wholly inside as one stored sum each, and only the fine cells along its edge are read and counted.

  1. Trace the polygon once: for each 100 m row, the runs of columns whose centres are inside.
  2. 10 km cells wholly inside add their stored sum.
  3. Inside the 10 km cells the edge cuts, 1 km cells wholly inside add theirs.
  4. Only inside the 1 km cells the edge still cuts are 100 m cells read and counted.

The answer equals a cell-by-cell count: checked against an independent Python count (rasterio, same rule) on Nairobi, Addis Ababa, a box across the Ethiopia–Kenya border, a square with a hole and all of Ethiopia, identical in population and cells; and for all 42 age and sex groups, within 7e-15 (float64 rounding). The response's method shows how many stored sums and 100 m tiles were used.

Sex and age. One 43-band file per country and year puts a cell's whole breakdown together, so one read returns every group. WorldPop's both-sexes files are not kept (each equals female + male within 1.5e-5 of a person) nor its 1 km mosaics (the sums here are exact); under 18 is kept, since five-year groups split 15–19. Against the total, the 40 groups agree within 3.1e-5 of a person per cell and about one person per country (Ethiopia and Kenya, all years).

One file per country and year holds 43 values per cell side by side: population, the 40 groups and under 18 by sex. Counting Ethiopia's whole border with every group took 39.4 s of CPU with 128 by 128 tiles and the floating-point predictor, and 5.6 s after the changes described below, against 3.7 s for population alone.

Several years. Which 10 km and 1 km cells lie wholly inside a polygon, and which 1 km cells its edge cuts, depends only on the shape, since every year's pyramid has the same grid. With years that plan is made once; then each year's stored sums and edge cells are read, four years at a time (two for a country-sized edge), the edge from that year's own files. Each year's answer equals that year asked alone, to the person: checked on the five test polygons for all 16 years (5 November 2026). Admin totals hold every year in one row, so years there costs nothing more.

Indicators are arithmetic on the 42 age and sex sums of the same cells, so they follow the counting rule. The names and definitions are those of WorldPop's API v3, so the two can be compared field by field: median age by linear interpolation inside the five-year group holding the half-way person (90 and over taken as 90–100); ratios per 100; shares of everyone with age and sex data. On the central Addis box, 2025, every indicator agrees with WorldPop's to its rounding (below).

Catchments. Each radius is a geodesic circle, built as a polygon of 256 points at that distance on a sphere of radius 6,371,008.8 m (WGS84's mean), then counted like any polygon. The 256-sided polygon lies inside the circle and covers 99.99% of its area. Five kilometres around central Nairobi, 2025: 935,402 people; a 64-sided circle gives 931,744.

Where the code runs. A Cloudflare Worker: JavaScript in a V8 isolate, no GDAL. It reads byte ranges from storage, inflates tiles with the runtime's native zlib and adds up. A file's tile index is read in pages, only where a polygon needs it: the age and sex files of countries that cross the antimeridian span the whole grid in 32 × 32 tiles (the USA's has 26.5 million), and loading such an index whole broke the Worker's 128 MB memory limit until 4 November 2026. Workers bill CPU time, not waiting; ours is pinned beside the bucket so its few hundred reads per polygon are local. Reprojection, clipped GeoTIFFs and bulk ranking would need a separate compute service; none is needed for these endpoints.

Speed

Measured 23 October 2026 from Helsinki on Ethiopia and Kenya data; "warm" is a second request on an instance holding the files' headers.

Polygon 10 km sums 1 km sums 100 m tiles CPU Answer
Box across the Ethiopia–Kenya border 1,944 0 0 22 ms 0.6 s
Square with a hole, Addis Ababa 33 1,560 0 35 ms 1.2 s
Addis Ababa 1 473 2 89 ms 1.3 s
Nairobi 2 511 4 112 ms 2.0 s, warm 1.3 s
All of Ethiopia 12,969 33,364 155 2.3 s 8.0 s, warm 4.4 s

With every age and sex group (groups=all), through wrangler dev --remote, so every read crossed the Atlantic; CPU measured locally, an upper bound:

Polygon 100 m tiles (32 × 32) Reads Bytes read CPU Answer
Addis Ababa 47 24 10 MB 0.6 s 1.3 s, repeat 0.3 s
Nairobi 66 23 11 MB 0.4 s 3.1 s, repeat 0.3 s
All of Ethiopia 3,079 558 51 MB 5.6 s 13.5–16.3 s

Live on 4 November 2026, after the memory fix, all of Ethiopia with every group answered in 25 s (9.4 s of CPU); city boxes in 0.5–1.4 s.

For an administrative unit, totals with groups answers in 0.2 s warm (Kenya's 47 counties, 1.3 s cold).

The shared cache. Requests rarely land on an instance that already holds a file's header: of ten calls in a row, most started fresh. Every instance in a Cloudflare data centre shares that data centre's cache, so the headers, tile indexes, VRTs and Parquet footers, and the 100 m tiles of city-sized shapes, are kept there after the first read; published files never change, so nothing kept can go stale. A country-sized edge skips it for tiles: filling the cache with Ethiopia's border over nine years cost 47 s of CPU against 31 s. Each answer's x-cache-reads header, and method.cache in polygon and catchment answers, count what came from the cache. Median of ten calls, server time, 5 November 2026, 2025 data:

Polygon No shared cache Headers and indexes cached Tiles cached too
Box over central Addis Ababa 446 ms 265 ms 59 ms
Nairobi county 451 ms 542 ms 132 ms
All of Ethiopia 3.2 s 3.5 s 3.2–4.6 s (tiles not cached; decoding is the cost)

Several years, live on 5 November 2026: the central Addis box for 2015, 2020, 2025 and 2030 in one request, 0.97 s on first read and 0.1–0.3 s after (WorldPop v3: 58 ms); Nairobi county for those four years with every group and indicators, 4.5–5.2 s; within 1, 5, 10, 25 and 50 km of central Lagos for all 16 years, 11–17 s; all of Ethiopia for nine years, 37–74 s.

Reading files directly from Helsinki:

Read Time
One year, a city window, from a COG (cold) 0.3–0.9 s, in 4–5 requests
Addis Ababa box, all 16 years, from the Zarr cube 2.4 s
The same, from 16 COG reads 11.0 s
Nairobi box, all 16 years, cube vs 16 COGs 2.0 s vs 4.7 s

Reading 2015–2030: sixteen COG reads, one per year, against one Zarr read where every chunk holds all sixteen years. Addis Ababa 11.0 s against 2.4 s; Nairobi 4.7 s against 2.0 s.

One year: a COG. A time series: the cube.

How it is built

The pieces: WorldPop's GeoTIFFs are downloaded, converted and checked on an ingest server, uploaded to R2 storage, and read either directly by range or through a Worker API that runs beside the storage.

One server in Helsinki (64 cores, 314 GB RAM) is the only writer. Per year it downloads every country, converts, checks, uploads, then builds the global VRT and the sum pyramid, each uploaded only if its checks pass:

What we learned

Limits

Licence and citation

WorldPop's data is licensed under CC BY 4.0 by WorldPop's licence statement. Cite the product each figure comes from, as WorldPop asks:

Changes made by data-geo: country files re-encoded as cloud-optimised GeoTIFFs with every value kept, age and sex files combined into one file per country and year, and new products computed from them: global VRTs, the Zarr cube, sum pyramids, admin totals and API counts. The checks are under How it is built.

WorldPop's notices, as they publish them: "The dataset currently represents an alpha version (R2025A) product and may change over the coming year as improvements are made." and "This data is not intended, and should not be used, for purposes that discriminate against, exploit, surveil, or otherwise harm individuals or vulnerable populations; users are expected to apply it responsibly, assess the risks of their own analyses (particularly when combining it with other datasets) and are solely responsible for how they use it."

Boundaries come from geoBoundaries; each unit's answer carries its own source, year and licence.

This page is rendered from /worldpop/how-it-works.md.