> Posted by Jeffrey Riecke, Communications Specialist, CFI
Going door-to-door to conduct surveys is expensive. Going door-to-door to conduct surveys assessing household consumption and poverty levels in far-flung areas around the world is even more expensive. And reliable data, of course, is crucial to financial inclusion and other international development efforts.
In recent years, the use of nighttime satellite imagery capturing civilizations’ lights or lack thereof has risen as a means to learn more about an area’s poverty levels without cumbersome surveys. But with these images alone, the picture is incomplete. A new project from a research team at Stanford University devised a computer model that brings poverty assessment into sharper focus. The model accurately predicts poverty levels, an ability built through machine learning using nighttime satellite imagery, high-resolution daytime satellite imagery, and household survey data. In fact, the model is able to predict up to 75 percent of the variation in local-level economic outcomes, and beats the nightlight models nearly all the time.
How does the model work and what are its limitations?
The model assesses physical features found in images to determine the poverty levels of the corresponding areas. But instead of simply looking at “light” versus “dark” areas, as in nighttime approaches, the model harnesses features in high-resolution daytime images, like paved roads; metal versus non-metal roofs; farmland; waterways; and urban marketplaces. Nighttime images might be effective in determining the levels of moderate to high wealth, but in impoverished areas, the nighttime images are just dark. For example, a village near a lake and a village near a forest might look equally dark in night images, but such villages would have different access to natural resources and this would have different effects on wealth.
The Stanford team used machine learning (the science of designing computer algorithms that learn from data) to create the model’s predictive ability. First, the researchers fed the computer model corresponding day and night images, using deep learning techniques to train it to predict where night lights would be by looking at daytime images. Second, once the model knew the connections between land features and lights, the researchers fed it actual survey data from the Demographic Health Services and World Bank Living Standards Measurement Study to further sharpen its ability to link physical features with poverty. Without being told what features to look for, the model is able to summarize huge amounts of information and recognize patterns that the researchers can’t, creating more accurate and more efficient estimates of where impoverished people live.
There are some important considerations concerning the limitations of the model. For this effort, the researchers looked specifically at Nigeria, Tanzania, Uganda, Rwanda, and Malawi. Eventually, they want to scale up the project to cover all of sub-Saharan Africa and in time the entire developing world. Poverty exhibited in the five countries studied for this iteration of the model might look different than poverty elsewhere in the region or beyond. Additionally, assessing poverty in rural versus densely populated urban areas presents unique challenges. Ultimately, the goal is to use the model in areas where surveys do not exist.
The new model and its approach are promising for poverty assessment. Machine learning is strongest when data is abundant, and both night and daytime satellite images and LSMS data is abundant and publicly available. We’ll look forward to following the advances of this and other computer-aided data collection approaches, both to determine the impacts of previous poverty reduction efforts, and to direct future efforts.
Image credit: Neal Jean
Have you read?