The Issue

Countries in emerging economies generally do not have reliable civil/health records, nor periodic, extensive or updated population and housing censuses. This can mean millions of children are left out and forgotten - unable to access the healthcare, education and protection they require. We want to make these invisible children visible.

Why Does it Matter?

Article 24 of the Convention on the Rights of the Child states that every child has the right to the best possible health. Similarly, the UN Sustainable Development Goals have a specific focus to ensure healthy lives and the promotion of well-being for all ages. However, with unreliable census data and health records, it is difficult to maintain an accurate understanding of the number of children in each community. This can heavily impact service planning and delivery, leaving some children at risk of missing out on life-saving vaccinations and treatment.

Estimates of population density are a valuable data source for policymakers, for example, in the context of developing a vaccination programme. However, census data is not collected frequently enough to be sufficient for such applications. This means that many census based population estimates do not account for rapid population change and may also lack precision in rural areas.

Given that populations can differ locally and change rapidly, it is essential to develop methods which can better assess population density frequently and with minimal resources. This will be vital in helping UNICEF to better tailor and target their vaccination programmes to reach the children most in need.

Our Project

This is the second phase of the Building Footprint Identification project. The objectives of this project are:

1. To use state-of-the-art machine learning tools to extract features from satellite images that are relevant for population density estimation such as building footprints and the distance to roads and markets

2. To explore statistical models that take into the consideration the spatial correlation of population density and compare them against models that ignore this assumption

3. To assess the applicability and effectiveness of end-to-end learning for estimating population density from satellite images in the two models above, and in the context of sparse survey data

4. To assess the sustainability of these approaches in terms of the resources required, e.g., amount of human supervision, resolution of satellite images

Our Outputs