Getting aid to impoverished Africans is hard enough, what with blockades of bureaucracy and red tape. But in many African countries, bad data, or a lack of it, makes distributing funds even more troublesome.
"Fighting poverty has always been this shining goal of the modern world," Neal Jean, a doctoral student in computer science at Stanford University's School of Engineering, told me. "It's the number one priority for the United Nation's 2030 Agenda for Sustainable Development, but the major challenge is that there's not enough reliable data. It's really hard to help impoverished people when you don't know where they are."
This fundamental problem was what Jean and five computer scientists hoped to solve, using satellite imagery and a machine learning model. Their new study, which was published today in Science, provides a proof-of-concept for an algorithm capable of predicting information about poverty in five African countries: Nigeria, Tanzania, Uganda, Malawi, and Rwanda.
Look at Angola, for instance. Forty years have passed since the country gained independence from Portugal, but its first postcolonial census was conducted just two years ago. The African nation is unfathomably rich in crude oil, but after 27 continuous years of civil war, half of its people live in poverty. Unfortunately, with scarce data on their economic well-being, it's nearly impossible to create programs that could help Angola's poorest communities, because no one knows exactly what is needed.
Countries can be loath to report their own inequality, due to corruption and conflict. According to the World Bank, 39 out of 59 African countries completed less than two population surveys related to poverty between 2000 and 2010. Of these nations, 14 reported no data at all, and most of the information amassed will never reach the public domain.
For decades, researchers have struggled to measure poverty using alternative data sets, such as social media, web search queries, and mobile network usage. In Rwanda, for example, where nearly 72 percent of people had mobile access in 2014, researchers were able to map their location based on the country's telecommunications data. While nontraditional methods like this were informative, the study mentions, they also raised issues of privacy and scalability, due their reliance on proprietary information.
Meanwhile, traditional collection efforts like household surveys were too expensive, costing hundreds of billions of dollars, and were sometimes hampered by civil unrest. Often, donors would offer African countries loans for census taking, instead of grants, which many could not afford to accept.But high-resolution satellite imagery—the kind you can find on Google Maps "satellite view"—is publicly available, geographically unlimited, and free. And as the study reveals, it can also hold a wealth of poverty data, if you just know how to look for it.
Using satellite imagery to estimate concentrations of wealth isn't new. Other researchers have already determined that "nightlights," or artificial light pollution, can be indicative of economic activity in developing nations. Theoretically, higher luminosity could mean more infrastructure, development, and wealth. Though it's worth noting that other models found it difficult to differentiate extremely low light levels from zero. For the same reason, nightlights are also less useful in places that are densely populated.
What's novel about the new study, however, is that the algorithm was able to close data gaps based on patterns it learned from looking at satellite shots. According to Jean, around 4,096 economic markers were identified in daylight images, including roads, urban areas, and waterways. By recognizing these features, the algorithm was then able to anticipate which areas would exhibit high luminosity at night.
"Our basic approach involved a machine learning technique called 'transfer learning,' which is the idea that you can solve a hard problem—in our case, predicting poverty from satellite images—by trying to solve an easier one. As you solve that, you can learn things that are transferable to help you solve the more difficult issue," Jean added.
Predicting light levels, instead of other wealth indicators, was important to the team because nightlights are among the most studied economic markers in developing nations. In the future, NGOs could potentially apply the algorithm to map poverty around the world. Compared to other methods, the machine learning model could also help to bolster transparency, since it doesn't rely on any proprietary data sets.
Still, the technique has its limitations, such as the unreliable frequency with which Google Maps updates its images. And the algorithm isn't perfect. According to the study, while it outperformed other data collection methods, the algorithm's predictions were only "fairly accurate."
There's also something fatally flawed about finding our poorest people by their glimmers of existence, only detectable from space—so unimportant to their nation's leaders or global community that the task of saving them is relegated to a machine.
Jean said he eventually plans to make their research open source, and is talking to several organizations about testing the algorithm in practical applications. "If we could provide them with high-resolution poverty maps, they could overlay them on regions where operations already exist, and ultimately inform where they distribute funding."
In 2012, the World Bank estimated that 330 million Africans currently live below the poverty level, surviving on $1.90 per day. Despite advances in technology and humanitarian efforts, the number of Africans living in extreme poverty has risen by almost 100 million since 1990.
"Better data will make for better decisions and better lives," said Luc Christiaensen, lead economist at the World Bank. "It is not just about quantity, the quality of the data also matters."