Skip to main content

<em>Science</em>: Predicting Poverty by Satellite

No video provider was found to handle the given URL. See the documentation for more information.

By combining satellite images and sophisticated machine learning, researchers have developed a new technique to estimate the poverty level of distinct villages in developing nations, according to a study featured in the 19 August issue of Science.

The technique combines daylight satellite images that capture features related to an area's average household income — such as the prevalence or absence of tin roofs and paved roads — along with images of nighttime lights, another important indicator of economic activity. The researchers designed an advanced computer program to identify such features in the images and to then use that information to estimate a village's wealth.

Governments and other international aid agencies need accurate data on poverty to guide spending on assistance projects, and a lack of data can hinder their efforts. In particular, data is limited in many African countries where assistance is in great need. According to the World Bank, 39 out of 59 African countries conducted fewer than two surveys between 2000 to 2010 substantial enough to result in measures to counter poverty.

"Household surveys are the traditional method for collecting these data, but these are expensive and time-consuming to conduct, and most developing countries do them rarely, if at all," explained Marshall Burke of Stanford University, a co-author of the study. "Our goal was to develop a cheap, scalable method for generating data on economic livelihoods in poor countries."

Some recent studies show that satellite data capturing lights visible at night can be used to predict wealth in a given area, where more lights indicate greater wealth. However, nightlight data alone are not effective at differentiating between the poorest regions. Satellite images appear uniformly dark, for example, between poor and ultra-poor regions that lie below the international poverty line of $1.90 per person per day.

To circumvent this problem, Burke and colleagues used a combination of nighttime and daytime satellite data. The daylight data capture features such as paved roads and metal roofs, markers that can help distinguish poor from ultra-poor villages. The team then developed a sophisticated computer learning algorithm to pore over the satellite images in search of such critical features.

"The nightlights help the algorithm figure out what features in the daytime imagery are predictive of economic activity," explained Burke. "The algorithm picks out things like roads, urban areas, farmland, and waterways, without being told to look for any of those features. In the last step of the process, we use these features in the daytime imagery to predict village-level wealth."

"The approach is surprisingly accurate, explaining over half of the variation in wealth that we observe in household surveys," he said.

The new model, tested in Nigeria, Tanzania, Uganda, Malawi, and Rwanda, outperforms nightlight-only models by 81% in predicting regions where household wealth falls under the international poverty line, and by 99% in areas where household wealth is two times below the poverty line. Importantly, the new method uses publicly available satellite data from Google Static Maps API and the process can be repeated more frequently than on-the-ground surveys. Furthermore, initial evidence suggests that a model "trained" in one country can be used in another.

Burke's team is studying whether lower-resolution imagery can also be used effectively to explain variation in economic activity. If it can, data collected over previous decades could be used to analyze changes in economic development over time.

"We [also] want to apply this set of computational tools to other social and environmental questions, such as predicting where armed conflicts break out, or understanding the determinants of agricultural productivity in different parts of the world," said Burke.

Burke was initially skeptical that this approach would work, noting that the application of machine learning in economics is still in its infancy. "But I'm pretty convinced now that these tools have a ton to offer the social sciences," he said, "and we've recently formed a lab at Stanford designed to get students and other faculty excited about using computational tools from computer science to solve pressing social and environmental problems."