Much like eradicating a disease, ending poverty requires knowing where it exists. A new project uses satellites and artificial intelligence to measure poverty rates – reaching places missed by traditional household surveys. Researchers say this can help track changes more quickly and improve the ability of governments and nonprofits as they try to end extreme poverty by 2030.
Poverty data is traditionally collected through household surveys. People go from home to home to ask a long list of questions regarding income, spending, access to schools and hospitals, assets and more. That information is used to determine poverty rates, maternal mortality rates and other well-being indicators.
Data in Africa is sparse. Between 2000 and 2010, 14 African countries did not conduct surveys. In fact, 39 of 59 countries conducted two or fewer surveys in the span of a decade. That means two-thirds of the continent does not have sufficient data for measuring poverty. The problem is that surveys are slow, costly and require coordination from federal to local governments. Data collection is getting better, but it is still not good enough.
“We are consistently limited in what kinds of projects you can look at, particularly in conflict-ridden countries,” Ariel BenYishay, economics professor at the College of William and Mary and chief economist for AidData, told Humanosphere. “The amount we are able to learn about foreign aid and whether it is working or not is continuously limited by the ability to send people out to ask people at their homes how they are living. To do that over and again is very costly and time-consuming.”
With AidData’s financial support, researchers at Stanford University are working to measure poverty with satellite images. Previous research showed that measuring night lights was a better way of estimating national level GDP. Access to electricity relative to population is a good, but imperfect proxy for development. The Stanford team went a step further by analyzing open access satellite images both at night and during the day.
Daytime images reveal more. Roads, farmland and waterways are visible and provide clues about the well-being in the region. The team trained an artificial intelligence program to learn from tracking nightlights and apply the same idea to daytime. They focused on parts of Nigeria, Tanzania, Uganda, Malawi and Rwanda where there is good data and large areas of people living in poverty.
The program began making predictions about poverty based solely on the images provided. It made decisions on its own to determine which households were below or above the poverty line. The predictions were then were compared against the survey results.
The researchers found that the daytime images were 81 percent more accurate at predicting households below the extreme poverty line ($1.90 per person) than nighttime ones. And even better at finding people three times below the poverty line.
“This is a cool technique that I’m sure will (and should) get a lot of use among researchers, and will hopefully catalyze further refinements to the approach,” said Justin Sandefur, a senior fellow with the think tank the Center for Global Development, in a blog post reacting to the findings.
Night lights are not all that great of a predictor of poverty. Access to electricity can mean a village has fewer people living in poverty. It could also mean an illegal network of wires were set up to provide electricity to some homes. The research team agrees.
“We should think of this first effort as a proof of concept and by no means the maximum accuracy,” said BenYishay. “The broader point that this is not the finished product we are going to hang our hats on.”
The program that the team set up is still running and learning. BenYishay said the refining process also allows for immediate learning about parts of countries where there are no surveys. If the promise of the findings become reality, researchers can start tracking development progress and inform policymakers.
“What we are really trying to do is learn about poverty in an area roughly the size of a village,” he said. “I think it will help answer questions about foreign aid impact in specific contexts. Policymakers could use it to improve their planning and see what is effective or not.”
Sandefur is more pessimistic. He sees the value of the information for researchers, but worries that the accuracy may not be up to snuff for policymakers.
“Contrary to popular belief, policy applications often require much greater rigor than required for publishing in academic journals,” he wrote.
BenYishay agrees that the current model is not accurate enough. But he is hopeful that it can keep learning. It may not completely replace traditional household surveys, but it could supplement available data and fill in the many information gaps across developing countries.
For AidData it can also be an accountability tool. If there is plan to install a water point for a village the images can track whether it exists and potentially if it is used. Over a region or entire country, that kind of information can help understand changes over time and better inform the global fight to end extreme poverty.
