U.S. Markets open in 6 hrs 12 mins

Machine learning ‘poverty map’ could help aid get to the right places in Africa

Luke Dormehl
Researchers at Stanford University are using machine learning and satellite data to develop a detailed “poverty map” of Africa by looking at areas where there there are few nighttime lights.

There are few bigger challenges than trying to solve world poverty. While there are plenty of initiatives going on in this area, one of the most intriguing is being carried out by researchers at Stanford University. Using a combination of satellite data and machine learning, they’ve developed a “poverty map” of Africa that could help direct aid to some of the world’s most deprived areas.

“One part of the problem when it comes to dealing with poverty is that we don’t have very good data,” Neal Jean, a Ph.D student in Machine Learning at Stanford, told Digital Trends. “If we want to help people, but we don’t know exactly where they are, that makes it very difficult to do. Traditionally, the way data is collected on poverty is by going out into the field and having people conduct surveys. But that’s a very slow process, incredibly human work-intensive, and not particularly scalable. Our objective in doing this project was to come up with a cost-effective and scalable way of filling in some of these data gaps.”

The idea of using satellite images for the work came about as a way of dealing with these so-called “data gaps.” Although there is a dearth of local survey information about poverty levels in individual African villages, there are plenty of satellite images. The task was therefore to come up with a way to use these images to extract valuable insights.

Related: When doctors can’t reach sick people in Madagascar, they send this medicine-carrying drone

The concept the researchers came up with was training a deep learning neural network with both day and night satellite images. In doing so, it was possible to identify where settlements existed, but where there were few nighttime lights — an observation that correlated with areas of impoverishment. The machine learning system the team used was eventually able to come up with a list of 4,096 features it could use to look at a particular area on the map and predict its level of nighttime lights.

Jean said there are two ways in which the project is progressing. “Firstly, we’re trying to expand the coverage of our study,” he noted. “Right now, we’ve only done it in five African countries, all of which are relatively similar visually. We’d be curious to see how this would work in other developing countries like India — or even in developed countries like the United States.”

The second development is coming from the various nonprofits and other organizations that hope to use Stanford’s poverty map to help them better distribute aid in Africa. “We know that these maps aren’t perfect yet, but hopefully they’ll be able to help guide some of the decisions made in terms of expanding focus in this area,” Jean concluded.

A new paper on the work was published in the journal Science,