video: This is a short animated video providing an overview of how satellite imagery and machine learning can be combined to map poverty around the world. This material relates to a paper that appeared in the Aug. 19, 2016, issue of Science, published by AAAS. The paper, by Neal Jean at Stanford University in Stanford, CA, and colleagues was titled, "Combining satellite imagery and machine learning to predict poverty." view more
Credit: Data from: N. Jean, M. Burke, M. Xie, W.M. Davis, D. Lobell, S. Ermon., "Combining satellite imagery and machine learning to predict poverty." Science, 2016
By combining satellite data and sophisticated machine learning, researchers have developed a technique to estimate household consumption and income. Such data is particularly difficult to obtain in poorer countries, yet it is critical for informing research and policy, and for efforts including resource allocation and targeted intervention in these developing nations. The African continent provides a particularly striking example of limited insights into economic wellbeing. According to World Bank data from 2000 to 2010, 39 out of 59 African countries conducted less than two surveys substantial enough to result in poverty measures. Surveys are costly, infrequent, and cannot always reach countries or regions within countries, for instance, due to armed conflict. Recent studies show that satellite data capturing nightlights can be used to predict wealth in a given area; however, nightlight data alone is not effective at differentiating between regions at the bottom end of the income distribution, where satellite images appear uniformly dark. To circumvent this problem, Neal Jean et al. turned their attention to daylight imagery, which offers higher resolution and can capture features such as paved roads and metal roofs, markers that can help distinguish poor and ultra-poor regions. The researchers then developed a sophisticated learning algorithm that categorizes these features. Several different validation methods reveal a high level of accuracy in their approach. The new model outperforms nightlight models by 81% in predicting poverty in regions under the poverty line, the researchers say, and by 99% in areas that are two times below the poverty line. Importantly, the new method uses publicly available daytime satellite data, can be repeated more frequently than surveys, and is inexpensive to use. Furthermore, initial evidence suggests that a model "trained" in one country can be used in another. In a related Perspective, Joshua Blumenstock discusses recent ways in which researchers have attempted to pinpoint impoverished areas, and how developments by Jean et al. offer a more accurate means to narrow in on the neighborhoods that would benefit most from social programs.
###
Journal
Science