Organizations often conduct door-to-door surveys to identify people living in poverty, but the downside is that these surveys are often time-consuming and expensive.
Indeed, locating impoverished environments is still a challenging process for researchers, and the availability of accurate information is still lacking.
Now, in a new study, scientists from Stanford University propose a more reliable method to map poverty in areas previously void of data — by combining satellite images and making use of machine learning.
Novel Method
Led by Stanford computer science doctoral student Neal Jean, researchers sought to determine whether the combination of high-satellite imagery and machine learning — the science of designing algorithms that learn from data — could predict estimates of areas where impoverished people lived.
Specifically, they extracted information about poverty from these satellite images, and built upon previous machine learning algorithms to detect impoverished areas across five countries in Africa.
What's challenging was that although standard machine learning can work when they access large chunks of data, in this case, there was little data about poverty to begin with.
Jean says there are a few places in the world where the computer can tell whether the people were poor or rich, making it difficult to extract information from the amount of available daytime satellite imagery.
Nightlight Data
Areas shown in satellite imagery that are brighter at night are usually more developed, researchers say.
The solution then produced another combination: putting together high-quality daytime satellite imagery and images of Earth at night. These are called "nightlight" data.
Jean and colleagues used the nightlight data to detect features in the higher-resolution imagery that are linked to economic development.
Jean says that without being told what to look for, the machine learned to pick out many things that are recognizable to humans from the imagery, including urban areas, farmland and roads. These features were then used to predict wealth in the village-level, as measured in available survey data.
In the end, Stanford scientists discovered that this method was efficient enough in predicting poverty distribution, and even outperformed existing methods.
The improved poverty maps can help policymakers and organizations distribute aid and funds more efficiently, as well as evaluate and enact policies more effectively.
"Our paper demonstrates the power of machine learning in this context," says Stefano Ermont, co-author of the study and a computer science assistant professor.
Ermont says that since the new method is cheap and scalable, it could be applied to map poverty across the world in a low-cost way.
Details of the new report are published in the journal Science.
Watch the video below.
