airmosaic-ai

WorldPop: population surfaces and conservation of counts

English 中文

Population grids are spatial estimates, not cell-by-cell enumeration. Choose the product and aggregation design to match the exposure question. Acquisition is described by the WorldPop skill.

Provenance and product selection

WorldPop teams at the University of Southampton and partners combine censuses, official estimates/projections, boundaries and geographical covariates. Underlying population and boundary inputs differ by country, year and release. See the official FAQ.

Product Selection implications
Global1 2000–2020 Distinguish constrained/unconstrained distribution and UN-adjusted/other total benchmarks
Global2 2015–2030 The reviewed R2025A v1 family includes country 100 m/1 km and global 1 km products; assumptions and release differ from Global1
Global2 country 1 km R2025A v1 People per pixel, constrained; the historical small-file example was LIE, 2020, UN-adjusted, not China or the full world
Catalog REST API Metadata discovery at https://hub.worldpop.org/rest/data; separate from scientific aggregation
Regional statistics API v2 Geometry-based statistics, a different service from raster discovery/download

Product snapshots above were reviewed on 2026-10-26. Confirm a chosen release record before new acquisition. Retain its license, recommended citation and DOI; the reviewed Global2 population product DOI is 10.5258/SOTON/WP00840, with CC BY 4.0 stated by its provider. Fix alias, ISO3, population year popyear, resolution, constraint, adjustment and release; publication dates do not replace population reference years.

Uncertainty

Census omissions, dates, boundaries, interpolation and projection affect totals and spatial allocation. Building/built-area errors affect constrained distributions; zero allocation is not a field-confirmed absence of people. Matching a national benchmark does not prove cell accuracy. Annual population does not describe hourly mobility or individual activity.

Global1 and Global2 may differ in inputs, methods and grids; raw subtraction/splicing can confound product changes with demographic change. The R2025A v1 release statement describes an alpha release subject to updates. It aligns country totals with WPP 2024 January 1 estimates; do not silently equate these to midyear population. Retain acquisition date and checksum.

Normalization and aggregation

  1. Select total population or age/sex groups and verify country, popyear, product, release and file identity.
  2. Decode the raster; check bands, CRS, affine/grid, pixel size, dtype, NoData, extent and valid values. Preserve metadata and checksums.
  3. Distinguish counts from density. People per pixel can be summed over compatible complete cells; people per hectare/km² require actual areas. Angular degrees are not km².
  4. Preserve NoData separately from zero. Require finite nonnegative valid values, investigate anomalies and avoid cellwise rounding that changes totals.
  5. Reproject/aggregate counts with an explicit mass-conserving allocation. Bilinear interpolation alone does not conserve population. Report target grid, area assumptions and before/after totals; densities and age shares need different treatment.
  6. For zonal sums, reconcile boundary vintage, CRS, grid overlap and coastal NoData. State partial-cell allocation rules and report incomplete regional coverage.
  7. Require mutually exclusive, well-defined age/sex groups; compare subgroup sums to totals in the same product/year. Align dates and denominators with GBD, concentrations and risk functions; total-population rasters alone do not supply all health inputs.

See shared data quality and population/health alignment.

This site is open source. Improve this page.