You might be thinking that the state of global poverty isn’t exactly a secret. Organizations like the IMF, for example, publish spreadsheets of numbers about economics development. But often there is a lot of guesswork behind these figures. Researchers gather this data by going door to door, and then extrapolating those figures out to a national level. Poor countries often lack the resources to do this; other times, governments just don’t want to, for fear of publishing statistics that show they are doing a bad job.
This is especially a problem in the parts of the world that suffer from the worst poverty, like sub-Saharan Africa. In fact, between 2000 and 2010, 14 African countries carried out no surveys from which national poverty statistics could be constructed at all, the World Bank says.
These statistics are important because international organizations, governments and charities need them in order to design poverty alleviation programs that work — for example, to meet the U.N.'s ambitious goal to end global poverty by 2030.
Neal Jean, one of the authors of a new study published in Science, says the U.N.'s goal will be very hard to accomplish without good measurement. “People are coming up with potential policies and interventions. But if you can’t measure your progress toward your goals, how do you know these policies are even effective?”
But satellite imagery offers a potential answer to this problem. And that answer is getting better and better with each passing year, as high-tech start-ups like Planet Labs and BlackSky Global launch new fleets of small commercial satellites that dramatically lower the cost and raise the availability of images from space. Researchers have used this kind of imagery to answer economic questions like how many people are visiting Chipotle, whether China’s real estate market is crashing or whether mines are producing at their full potential.
Other researchers have measured economic development by looking at lights at night, since the places that appear the brightest are typically the most developed — areas with streetlights, businesses and other infrastructure. In 2011, economists at Brown used night lights from space to show economic fluctuations, like how Indonesia’s economy slowed during the Asian financial crisis. In 2014, economists at the Federal Reserve Bank of New York used a similar technique to measure India’s growth.
In a paper newly published in the journal Science, Jean and other researchers have echoed these techniques and expanded on them in an important way. They feed images of the Earth during the day and the night through a machine learning program to calculate how much households own and consume in five African countries — Nigeria, Tanzania, Uganda, Malawi and Rwanda.
The main problem is that, while there are a lot of satellite images out there, there aren’t many that are accurately labeled as showing people that are rich or poor. To learn to predict poverty accurately from satellite images, a computer algorithm would need large amounts of labeled data like that to train on.
The researchers bypass this by using a technique called “transfer learning.” Transfer learning is the idea that you can solve a hard task by first solving an easy task and then applying what you’ve learned, says Jean. “The hard task that we actually care about is predicting poverty from satellite images, and the easy task is using satellite images in the day to predict how bright an area is at night.”
So as a first step, they train the computer to learn to predict how bright the lights will be at night by looking at satellite images taken during in the day. To do that, the computer’s image recognition system learns to look for things in the daytime images that are correlated with economic development — like roads and farmland, even the size and material roofs are made of, and the distance between villages and cities.
The researchers don’t tell the computer program what to look for — it just identifies correlations on its own. As a second step, the researchers then have the algorithm use those features from the daytime imagery to predict wealth and consumption at the village or community level.
In general, the researchers are able to predict about 50-70 percent in the variation in wealth using just the daytime satellite images. That may not sound like much, but they say that it’s actually far more accurate than any existing approach out there. In addition, their method is very low cost, and it should continue to gain in accuracy as satellite imagery improves.
“The bar is very low,” Jean says. “Any new technique that we can provide that does better is going to be helpful for the people who are working on solving these problems.”