Skip to main content Skip to secondary navigation
Main content start

Satellites enhance sustainable development inquiries

A new paper in Science by King Center researchers documents progress in using satellite imagery and machine learning techniques to study development trends.
satellite image of a border town
Satellite image of a border town | Credit: iStock
Innovations in Methods and Data

Several years ago, when a trio of Stanford University professors first proposed combining satellite imagery and machine learning approaches to study development trends—including poverty levels, agriculture outputs, and population density—potential supporters remained skeptical.

A lot has changed since then.

In 2017, the Stanford King Center on Global Development, then named the Center on Global Poverty and Development, provided financial support to those faculty—Earth System Science Professors Marshall Burke and David Lobell, and Computer Science Professor Stefano Ermon—and others by launching the Data for Development initiative, from which some of the earliest research showing the method could work emerged. And, in the past five years, as more satellites have been sent into orbit, beaming higher and higher resolution images back to Earth, additional social and computer scientists have taken up the effort to put those images to use in the public interest.

In a new paper out in Science magazine, Burke, Lobell, Ermon, and Center on Food Security and the Environment research data analyst Anne Driscoll set out to see what progress has been made.

Marshall Burke
Marshall Burke

“The field is moving so fast,” says lead author Burke. “We thought it lacked a comprehensive summary for a lot of the key outcomes.”

But the paper goes beyond summarization; it also quantifies progress.

“We wanted to try to understand: how well are these methods working overall?” Burke says. “Are they as good as the data we traditionally collect?  Are they getting better over time?”

The results are promising. In four areas closely linked to the United Nations’ Sustainable Development Goals—smallholder agriculture, economic livelihoods, population, and informal settlements—the team found that prediction capabilities are “reasonably strong and improving” because of a combination of higher-quality images from all those satellites, more “training data” (traditional survey data used to train and validate the machine learning process), and researchers’ creativity in putting those materials to use.

“For multiple outcomes of interest, satellite-based estimates can now equal or exceed the accuracy of traditional approaches to outcome measurement,” the authors write in the paper.

Lobell says part of the goal of the paper is to describe ways to measure machine learning performance for development outcomes.  

David Lobell
David Lobell

“A lot of machine learning has focused on tasks with very clear benchmarks, like picking out cats in photos,” Lobell elaborates. “There it’s fairly easy to define performance. But we are often dealing with outcomes that have very imperfect traditional measures. If I ask you how much money you spent last month, or how many tomatoes your garden produced last summer, you likely won’t know exactly.”

Burke knows firsthand what a difference the technology can make. The development economist worked for years on the ground in western Kenya, collecting data and trying to measure farmers’ productivity in the region by sending surveyors to local households to query their occupants.

“We were somewhat frustrated by our inability to do that on any sort of scale,” he says. But, after he, Lobell, and Ermon decided to pair machine learning, or artificial intelligence, techniques with satellite images to do the same thing, they were amazed by the results. 

“We found what we could learn from satellite images was as or more accurate than what we were collecting on the ground,” he says.

The Data for Development initiative team also has used satellite imagery to assess village-level wealth in five African countriespredict the presence of sewage and water pipes as a measure of clean water and sanitation; and compare the machine learning process’ accuracy to survey data on wealth for 20,000 African villages dating back to 2009, among other inquiries.

Burke, Ermon, and Lobell also developed a course—Data for Sustainable Development—that introduces Stanford students to the method by allowing them to work on real-world applications, and regularly bring in experts to discuss their own use of the technique, including an event last spring with representatives from Planet Labs, Orbital Insight, and Descartes Labs.

Besides providing additional proof of the accuracy and effectiveness of satellite-based estimates, the paper makes several other key points. For instance, in order for the field to progress further, the authors write, more traditional survey data is needed to train and validate machine-learning models.

stefano ermon
Stefano Ermon

“It’s a chicken and egg problem,” Ermon says. “It’s difficult to know if something works in a setting in which you don’t have that much data.”

The authors also write that machine learning will never eliminate the need for on-the-ground data.

“What you can see from a satellite is not everything,” Burke says. “We can see certain things about households, but we can’t, for example, see within households” to understand how children are doing or answer questions about gender empowerment.

Another finding: despite growing evidence of the model’s success, there have been relatively few efforts to use satellite-based estimates in the public-sector policy- and decision-making process.

Here, Burke, Lobell, and Ermon hope to be helpful again. In 2019, with funding from the Rockefeller Foundation, they launched Atlas AI, a public benefit corporation that aims to use existing and future data to help solve important societal problems. Earlier this year, they completed a project with the Ethiopian government that sought to better target social safety net programs to the people that need the most help.

Burke says the King Center’s Data for Development Initiative funding was “catalytic” in the field.

“I feel very lucky to be part of a team at the cutting edge of innovation,” he says.