Mapping Extreme Poverty With Satellites and AI

This article provides an overview of how humanitarian organizations and the United Nations utilize satellite daytime imagery, night-lights data, and artificial intelligence to identify extreme poverty and distribute cash aid. Historically, relief agencies relied on expensive and slow door-to-door surveys to locate vulnerable populations. Today, machine learning algorithms analyze nighttime illumination alongside high-resolution daytime images of infrastructure, roof materials, and farmland to create accurate poverty maps. This data-driven approach allows organizations to bypass logistical barriers and deliver financial assistance directly to those who need it most, faster and more cost-effectively than ever before.

The Traditional Challenge of Poverty Mapping

Before the integration of satellite technology, the United Nations and humanitarian groups depended on traditional census data and household surveys. Conducting these surveys in developing nations is expensive, time-consuming, and often dangerous due to conflict or lack of infrastructure. As a result, many countries go over a decade without updating their demographic data, leaving aid organizations essentially blind when crises strike.

Night-Lights as a Proxy for Economic Activity

To solve the data gap, researchers first turned to night-lights satellite data. Satellites capture the amount of artificial light emitted from the Earth's surface at night. Because electricity usage and outdoor lighting strongly correlate with economic development, bright areas typically indicate wealth, while dark areas suggest lower economic activity. However, night-lights data alone has limitations. It struggles to distinguish between an unpopulated forest, a modestly poor rural village, and a region experiencing extreme poverty, as all three appear entirely dark from space.

Adding Daytime Imagery and Transfer Learning

To overcome the limitations of night-lights, scientists introduced daytime satellite imagery combined with an artificial intelligence technique known as transfer learning. AI models are trained to compare nighttime data with daytime images of the exact same locations. By doing this, the AI learns which daytime visual features correlate with the wealth indicated by the night lights.

Once trained, the AI can look solely at daytime images of pitch-black regions and still predict poverty levels accurately. The algorithm scans for specific visual markers of wealth or poverty, including:

• Roofing materials: Differentiating between thatched roofs (often indicating poverty) and metal or tiled roofs (indicating higher income).
• Road quality: Identifying paved roads versus unpaved dirt tracks.
• Infrastructure: Locating proximity to urban centers, markets, and waterways.
• Agriculture: Assessing the size, geometric regularity, and health of surrounding farmland.

Targeting Cash Aid Distribution

Once the AI generates high-resolution poverty maps, the United Nations and non-governmental organizations (NGOs) use them to execute targeted cash transfer programs. Instead of shipping physical goods like food and clothing—which can disrupt local markets and incur massive transport costs—organizations send money directly to individuals, often via mobile phones.

A prominent example of this occurred in Togo during the COVID-19 pandemic. The Togolese government, working with international researchers and humanitarian groups, needed to distribute emergency cash to its poorest citizens quickly, despite lacking recent census data. They used AI and satellite imagery to analyze the entire country and identify the poorest cantons. They then combined this spatial data with mobile phone metadata to pinpoint impoverished individuals and instantly transfer funds via mobile money networks.

Similarly, the United Nations Development Programme (UNDP) and the World Bank use these AI-generated maps to allocate resources, plan infrastructure projects, and ensure that emergency financial aid reaches marginalized communities that were previously invisible on traditional maps.