Harnessing machine learning and big data to fight hunger
By Melanie Lefkowitz
A group of Cornell researchers has received a $1 million grant from the U.S. Agency for International Development to use machine learning to rapidly analyze agricultural and food market conditions, aiming to better predict poverty and undernutrition in some of the world’s poorest regions.
The method will use open-source, freely available satellite data to measure solar-induced chlorophyll fluorescence (SIF) – photons emitted from plants during the process of photosynthesis, detected by satellite, which can monitor agricultural productivity. It will also consider land-surface temperature, which provides information about crop stress under water deficit or excessive heat, as well as food-price data.
“A method that can use near real-time, low-cost or freely available remotely sensed data can speed up the delivery of this information, and sharply reduce the cost,” said Chris Barrett, the Stephen B. and Janice G. Ashley Professor of Applied Economics and Management in the Charles H. Dyson School of Applied Economics and Management, and the principal investigator on the three-year grant. “If you are a humanitarian organization trying to really target your resources at the poorest rural areas, this seems a powerful diagnostic tool.”
Measuring agricultural health is essential in assessing conditions in poor, rural regions where the economy relies heavily on farming and where droughts, floods or crop failures can be devastating. Quickly identifying potential crises can help direct aid where it’s most needed. It can also guide farmers to plant drought-resistant or shorter-cycle crops, depending on conditions.
Traditionally, governments and aid organizations have used household surveys to locate at-risk areas, but surveys are expensive and time-consuming. A typical survey might take two years to conduct and analyze, so using data that can be observed and assessed in near-real-time can help deliver more timely and better-targeted aid. For example, children whose mothers were pregnant during times of nutritional stress are more likely to suffer lifelong problems, so delaying help to these families could have a permanent impact.
“These relatively seasonal fluctuations in well-being, in very poor communities where there isn’t a strong safety net, wind up having long-term consequences,” said Barrett, who is also a professor in the Department of Economics and a fellow of the Atkinson Center for a Sustainable Future.
Machine learning is a kind of artificial intelligence in which computers use uploaded data to train themselves how to solve specific problems. In this project, the researchers will use household survey data to train a machine learning model which will consider SIF, land-surface temperature and food-price information to predict where extreme poverty and undernutrition are likely to strike.
“We have some really big data that’s too complex to use directly, so we apply machine learning methods to extract important features. These features may be smaller in number but have higher predictive value than the surveys,” said David Matteson, associate professor of statistical science in Computing and Information Science and the School of Industrial and Labor Relations. “Our hope is that this higher predictive power will allow interventions to be closer to real time.”
The project builds on recent work led by Ying Sun, assistant professor of geospatial sciences in the School of Integrative Plant Science, Matteson and Yanyan Liu, senior research fellow at the International Food Policy Research Institute and an adjunct professor at the Dyson School. They received a grant last year from the Atkinson Center’s Academic Venture Fund to explore whether SIF data could predict areas of extreme poverty.
In addition to SIF, the machine learning model will consider land-surface temperature and food-price data. Though sensitive to weather conditions, land-surface temperature captures surface moisture and plant-water content in nearly real time, providing valuable information for monitoring crop health or drought risk. Adding food prices, which affects how much farmers earn and how much consumers can afford to buy to feed their families, should add value to the model, the researchers said.
From these data, the model will generate maps showing factors like estimated prevalence of various poverty rates or populations of children or women at risk of malnutrition. The maps will not only make it easy to identify specific regions needing help, they will also show how conditions evolve over time to help policymakers or aid organizations make decisions.
Once they’ve successfully built the model, the research team will offer training so governments and aid groups can use these tools themselves.
“Right now, humanitarian organizations are routinely going in and drawing lines on maps to say we’re going to work in these communities, and they often don’t have especially strong data platforms for making those decisions,” Barrett said. “You also have lots of policies that have geographic boundaries to them, and you want to evaluate how effective those policies are at reducing poverty or improving child nutrition. These sorts of measures can give you the underlying data with which to do rigorous impact evaluation.”
The research team also includes Liu, Leiqiu Hu of the University of Alabama and Linden McBride, Ph.D. ‘18, of St. Mary's College of Maryland.