SEATTLE — Household-by-household surveys by local census bureaus used to be the only method to collect household statistics, making mapping poverty inefficient and costly. By teaching computers to analyze high-resolution satellite images, Stanford’s machine learning algorithm now makes predicting poverty quick and easy.
From small-area estimation, direct household surveys to more sophisticated methods like a multivariate weighted basic-needs index, current poverty mapping methodology depends heavily on firsthand survey data. The great cost of obtaining more data in order to perform a better analysis is hardly justifiable. The lack of data regarding infrastructure and services in poorer areas in the world then impedes policymakers’ ability to make well-targeted plans.
In August 2016, researchers at Stanford University officially revealed their reimagined method for poverty mapping with a paper published in Science. Utilizing the machine learning algorithm, computers can now calculate per capita consumption expenditure of a certain location when provided with its daytime and nighttime satellite imagery.
The Technology Explained
Machine learning, a subfield of artificial intelligence, focuses on the computational model that allows an algorithm to learn to analyze raw data without being told what to look for.
Led by assistant professor of earth system science Marshall Burke and electrical engineering Ph.D. student Neal Jean, the team at Stanford created a machine learning algorithm that crawls through the millions of available high-resolution satellite images of poorer regions in the world, looking for visual evidence of the economic situation of those areas. Stanford’s machine learning algorithm compares the presence of light in daytime and nighttime images of a region to predict its economy activity – a technique known as transfer learning.
Using nighttime imagery as a reference, the algorithm picks up on a well-lit area and crosschecks it with daytime imagery to confirm its infrastructure development. A brighter area at night usually means more activities that involve electricity, hence the region is wealthier.
Via the process of familiarizing itself with the appearance of more developed areas from millions of images, the algorithm learns to spot visual details like streets, highways, waterways, farmlands and urban areas. These visual details are then used as filters in the later estimation of poverty.
Before it makes a final judgment, the algorithm will crosscheck its visual observations with survey data to improve its accuracy. The algorithm then can generate a prediction of poverty distribution of the area based on all previous assessments.
What to Expect Next
The research performed poverty predictions for five African countries, including Nigeria, Uganda, Tanzania, Rwanda and Malawi. These countries all have relatively substantial ground-based survey data to verify the accuracy of the predictions generated by the algorithm.
Stefano Ermon, an assistant professor of computer science at Stanford and a member of the project, said that the village-level wealth predictions generated from the machine learning algorithm are “very close” to predictions calculated with field-collected data.
In an interview with Science, political scientist from Columbia University Marc Levy said Stanford’s machine learning algorithm is “vastly more powerful” than traditional surveying, but he also cast doubts on the algorithm’s compatibility with more urbanized regions in the world.
More people have realized the tremendous potential of machine learning technology. After Defense Secretary James Mattis expressed his envy for artificial intelligence used by tech companies, the Pentagon began offering a prize of $100,000 for an algorithm that can read satellite images in the same way that Stanford’s machine learning algorithm does.
Researchers at the University of Buffalo are also using machine learning and satellite imagery to map poverty. They also add cell phone records to their raw data sets, offering a new approach to predict poverty by analyzing communication activities.
Eliminating poverty has been the priority of the United Nations’ Sustainable Development Goals for 2030. Both working towards the goal and tracking the progress require more reliable and frequent data on poverty across the globe. Stanford’s machine learning algorithm, a revolutionary method for poverty mapping, might well play a pivotal role in the world’s collective efforts in eliminating poverty.
– Chaorong Wang