As much as that data has the potential to provide invaluable information to humanitarian organizations, watchdog groups, and policymakers, there is too much of it to sift through in order to draw insights that could influence important decisions. A team of researchers from Stanford University, however, says it has developed an efficient way. By creating a deep-learning algorithm that can recognize signs of poverty in satellite images – such as condition of roads – the team sorted through a million images to accurately identify economic conditions in five African countries, reported the scientists in the journal Science on Thursday.
For the majority of the world, we don’t have any labels for [satellite] images, so it’s not like people have gone and looked at satellite imagery and said, ‘Ok, here’s a house, here’s a tree, here’s a road,’” Neal Jean, a graduate student in electrical engineering at Stanford University and lead author on the Science paper, tells The Christian Science Monitor. “Since there’s so much imagery, a big part of the problem that we face...is figuring out how to extract useful information from this unstructured data.”
Science - Combining satellite imagery and machine learning to predict poverty
Nailing down how to do this would be a boon for international efforts to track poverty and take stock of general economic conditions around the globe. In some parts of the developing world, international aid organizations such as the World Bank are experimenting with using satellite surveys to collect data remotely, instead of in person, house by house -- a tactic that could save both time and money. In places where there is unreliable data or none available at all, such as in North Korea, satellite photos showing no lights in the country versus the illumination of the world around it, can provide the only insights into economic activity on the ground.
Accurate information about people's needs could influence decisions about where to send aid or build roads or hospitals. On a larger scale, such geographic specificity could help track whether global efforts to reduce poverty in some regions are paying off.
As Mr. Jean and his team point out in Science, “data gaps on the African continent are particularly constraining.” In the first decade of this century, 39 out of 59 African countries conducted fewer than two national surveys that could help paint a picture of poverty conditions there, according to the World Bank. Most of that data is not even publicly available. And 14 countries had no surveys at all.
“These shortcomings have prompted calls for a ‘data revolution’ to sharply scale up data collections efforts within Africa and elsewhere,” writes Jean and his co-authors.
In response, many efforts are underway to apply advanced technologies to poverty alleviation efforts. Orbital Insight, a Palo Alto satellite data analysis company, has worked with the World Bank to help the organization determine where it should allocate more than $100 billion worth of loans each year, as Bloomberg reports. The company uses machine learning to find clues of poverty in reams of photos. Its software counts cars, analyzes the height and shapes of buildings, and measures agricultural activity in remote villages.
“If you see more cars, or more cars over time, that could be an indicator of relative wealth in one village vs. another that hasn’t seen growth in cars over time,” Jeff Stein, vice president for business development at Orbital Insight told Bloomberg.
Joshua Blumenstock, a data scientist at University of California in Berkeley, uses proprietary mobile phone metadata in countries like Rwanda to analyze patterns of calls. Those patterns give clues to whether people in impoverished communities have jobs, among other indicators of wealth. And GiveDirectly, a New York City-based nonprofit that gives cash to poor people in countries including Kenya and Uganda, has experimented with mining satellite images to determine who should get donations. People living in houses with thatched roofs are more likely to be poor than those living in houses with metal roofs, the organization figures.
While these are all useful methods, Jean’s technique has some additional benefits. For one, his team used publicly available data. It included daytime satellite photos from Google, night-time satellite data from the National Oceanic and Atmospheric Administration, and survey data from the World Bank. Also, the team didn’t need to teach its algorithm to look for one or two specific features of poverty in the daytime photos. The machine identified a handful of features on its own by comparing daytime photos with night-time ones.
Night-time satellite images are considered a good way to pinpoint poverty based on how much light sparkles from a region. The more there is visible light, the richer the area must be is the assumption, which is why the pitch-black satellite images (see above) from North Korea – juxtaposed against the widely illuminated South Korea – are so stark.
Measuring consumption and wealth remotely
Nighttime lighting is a rough proxy for economic wealth, and nighttime maps of the world show that many developing countries are sparsely illuminated. Jean et al. combined nighttime maps with high-resolution daytime satellite images (see the Perspective by Blumenstock). With a bit of machine-learning wizardry, the combined images can be converted into accurate estimates of household consumption and assets, both of which are hard to measure in poorer countries. Furthermore, the night- and day-time data are publicly available and nonproprietary.
Reliable data on economic livelihoods remain scarce in the developing world, hampering efforts to study these outcomes and to design policies that improve them. Here we demonstrate an accurate, inexpensive, and scalable method for estimating consumption expenditure and asset wealth from high-resolution satellite imagery. Using survey and satellite data from five African countries—Nigeria, Tanzania, Uganda, Malawi, and Rwanda—we show how a convolutional neural network can be trained to identify image features that can explain up to 75% of the variation in local-level economic outcomes. Our method, which requires only publicly available data, could transform efforts to track and target poverty in developing countries. It also demonstrates how powerful machine learning techniques can be applied in a setting with limited training data, suggesting broad potential application across many scientific domains.
SOURCES -Journal Science, CS Monitor, Youtube