Using Deep Learning to Track Poverty with Satellites


Researchers at Stanford University are utilizing artificial intelligence (AI) to identify areas of poverty in hard-to-reach places. Publishing their research in Science Magazine, the team consisting of Neal Jean, Marshall Burke, Michael Xie, Matthew Davis, David Lobell, and Stefano Ermon used deep learning algorithms to sort through millions of satellite images to identify economic conditions in five African countries. This research supports a forecast Tractica made over a year ago in our Artificial Intelligence for Enterprise Applications report, that spending on AI software by philanthropy organizations will grow dramatically over the next 10 years.

Why There Is a Need

Traditionally, philanthropy organizations have conducted door-to-door surveys to identify people living in poverty, but these surveys are imprecise, time-consuming, and expensive. Many international aid organizations including the World Bank have been trying to use satellite surveys to gather data remotely on developing countries, but the expense of gathering and analyzing this data using conventional methods has proven prohibitive.

How It Works

Nighttime lighting can be used as a rough proxy for levels of economic development. Areas shown in satellite imagery that are brighter at night are usually wealthier and nighttime maps of the world show that many developing countries are illuminated only sparsely. The team at Stanford compared high-quality daytime satellite imagery with images of the Earth at night. Benchmarking against the door-to-door survey data from five African countries – Nigeria, Tanzania, Uganda, Malawi, and Rwanda – the Stanford team trained a convolutional neural network (CNN) using satellite data to identify image attributes that can explain up to 75% of the variation in local economies.


The method, which requires only publicly available data, holds the potential to transform efforts to track and target poverty in developing countries. If more accurate information about areas of poverty were available, it could influence decisions about where to send aid workers or where to build roads or hospitals. At a bigger level, accurate geographical data about poverty could be used to optimize global efforts to reduce it. The researchers did not need to code a program to look for specific features of poverty; instead, the deep learning algorithm was able to identify poverty from the data. The method also used publicly available data, including daytime satellite photos from Google, nighttime satellite data from the National Oceanic and Atmospheric Administration (NOAA), and survey data from the World Bank, an approach that minimized the cost.


Almost from the first research grant that promised to create artificial intelligence in a summer, AI has been over-sold and over-hyped as much as any other software technology. Especially in the case of an application that promises to do so much good, onlookers may be reluctant to question the results. For example, based on nighttime lighting analysis, it might be reasonable to conclude that Las Vegas is the wealthiest city in the United States, when it is, in fact, only seventh. AI is also controversial because it can upset human social hierarchies. Characterizing different groups of people based on their wealth seems inherently controversial and may be run into resistance in some areas and in some circles.


Pioneers like Jean, Burke, Xie, Davis, Lobell, and Ermon are part of the reason we feel the philanthropy sector is well poised to leverage artificial intelligence. In Tractica’s Artificial Intelligence for Enterprise Applications report, we forecast that spending on AI software in the philanthropy industry will grow from less than $1 million in 2015 to $27.7 million by 2024.

AIE-15 chart Philanthropies

Comments are closed.