Case study

How DFID used satellite images to estimate populations

Learn how the Department for International Development (DFID) used computer vision to estimate populations.

This guidance is part of a wider collection about using artificial intelligence (AI) in the public sector.

AI application used

  • computer vision

Objective

The Department for International Development (DFID) needed to help developing countries better understand their population distribution.

Situation

Developing countries need to understand their population distribution so they can plan services. This information is often recorded by conducting a conventional census, however this was not always possible if there was conflict or insecurity.

Action

DFID partnered with the University of Southampton, Columbia University and the United Nations Population Fund (UNFPA) to apply a random forest machine learning algorithm to satellite image and micro-census data. DFID analysed the satellite image data to identify features such as:

  • settlement boundaries
  • buildings
  • transport networks
  • waterways
  • lighting
  • industrial areas

The algorithm then used this information to predict the population density of an area. The model also used data from micro-censuses to validate its outputs and provide training data for the model.

Impact

DFID has deployed the programme in Nigeria, Zambia, Mozambique and Democratic Republic of Congo (DRC). DFID has also conducted scoping missions to Tanzania, Ethiopia, and South Sudan.

The model - called GRID3 - is being used to:

  • develop a hybrid census model which combines population estimates for small areas of uniform, detailed grids with modelled population estimates
  • support developing countries as they plan their full censuses
  • plan vaccination campaigns and other services
  • support developing countries such as DRC safely gather population estimates for areas in conflict

DFID is now exploring whether they can also use mobile phone data to provide more accurate local level population estimates.

Updates to this page

Published 10 June 2019