Measuring poverty through satellite images

Poverty research of today is dependent on survey data from interviews with people living under the conditions that are studied. Gathering statistics from rural areas of Africa, for instance, is both costly and a slow way of learning about the situation, leading to a lack of data far from being sufficient to get a good overview of the living conditions in Africa.

“Better knowledge about the living conditions would mean better tools in fighting poverty”, says Adel Daoud, poverty researcher and project leader of two projects funded by the Swedish Research Council and Formas, with the ambition to create an algorithm that can look at satellite images and tell us the status of both health and economic conditions of the population in the area. Adel is also an affiliated researcher at Chalmers Data Science and AI division.

Data from surveys and satellite images teaches the AI system how to detect poverty

The project is a collaboration between social scientists at the University of Gothenburg and computer scientists at Chalmers. Also, researchers from the Department of Statistics at Harvard and from the Institute for Analytical Sociology, Linköping University, will participate in the project. In the project, data from surveys and satellite images are linked together to teach the AI system how to detect different aspects of poverty. The algorithm compares images from 1984 up to 2020.

“The algorithm learns what is characteristic for places with high poverty when viewed from above. It predicts what type of fuel that is common, if cell phones are common, and if there are hospitals and access to education. What are the most common means of transportation, and do people in general have bank accounts in the area?”, says Fredrik Johansson, Assistant Professor at the Data Science and AI division, Chalmers.

Poverty traps will be studied later

A later part of the project deals with using the data to study so called poverty traps, where societies seem to be in a loop of poverty despite initiatives to rise from it. The AI system will provide data that may be used to evaluate factors that have impact on poverty and living conditions, like political decisions, infrastructural initiatives and more.

“Why are some governments better than others in fighting poverty? Are there political strategies that are more successful than others? Is there a railroad between villages that have improved living standards, or has the government in the country gone from an authoritarian regime to democracy?“ says Adel Daoud.

The prospects look good. Already, the algorithms have proven to be very efficient in supplying predictions of living standards.

“Our first results are very promising. In particular, we are excited to see that our models are able to predict poverty levels at different points in time than the ones they were trained on. This takes us closer to the goal of identifying poverty traps", says Fredrik Johansson.