URBANA, Ill. – When international aid organizations respond to hunger crises around the world, they rely on food security early warning and monitoring systems. However, assessments from those systems can lag months behind the actual situation on the ground.
More accurate and timely warnings could greatly improve the global response to food crisis and help alleviate acute problems, according to a team of agricultural economists at the University of Illinois, working in collaboration with the University of Texas, Austin.
The researchers developed a prediction model that uses readily available weather, market and demographic data to forecast village-level food security status. They trained and tested the model using data from Malawi, a country with frequent food insecurity problems. They were able to predict food security status with up to 99% accuracy for the most highly food insecure village clusters, says Hope Michelson, assistant professor of agricultural and consumer economics in the College of Agricultural, Consumer and Environmental Sciences at U of I, and co-author of the study.
“We’re addressing an important problem that many people don’t know exists, which is that the way food security crises are assessed now is infrequent and must rely on uneven data,” Michelson says. Current early warnings are generally based on information collected on the ground and evaluated by local advisory groups. The process can be slow, using coarse data, and is prone to accusations of politicization by national governments and organizations.
Availability of data for prediction has vastly improved over the last 20 years, Michelson points out, and integrating temporally and spatially richer data into prediction models can provide relief organizations with information that is much more timely, accurate, and objective.
For the study, the researchers used data from 2010-2011 to predict village food security in 2013, and then compared the forecasts with measured 2013 food security outcomes. They included satellite data on weather and precipitation, market price data, and demographic data that represent household resources and ability to cope with food price shocks.
The research team then used the information to estimate and forecast village food security using three different internationally recognized food security measures. When comparing their predictions with the actual food security status for Malawi households in 2013, they found that their results ranged in accuracy from 83% to 99%, depending on the type of food security measure. They compared their results with the ability of current assessments to predict the same villages’ food security and found that the current assessment method had an accuracy close to or below 10%, Michelson says.
The goal of the project is to provide national stakeholders and global aid organizations with a timely, precise, and transparent model they can use to quickly identify regions that are likely to experience hunger crises, Michelson says. This model can provide an important contribution to an early warning system that would allow for immediate action as soon as a problem is identified.
While more complex machine learning models might further improve prediction, the researchers opted to keep their model accessible for governments and organizations. In a current extension of their work, however, they do apply machine techniques.
“We want to be clear that we are not advocating for replacing the current system of early warning,” Michelson states. “Rather, we hope to complement current processes. Assessing and forecasting food crises is hard, important work. It is our intention to contribute to that ongoing effort; to focus attention on particular sub-national regions as their situation worsens in real-time, for example.”
The next step is to systematize the information and build a model that can run in real-time, providing constant updates on food security status. The National Center for Supercomputing Applications at U of I is collaborating on these extensions, including a public website that monitors circumstances in real-time. They will start with data from Malawi, then gradually expand to other countries, as well as longer time horizons.
One challenge will be to expand the model to areas where information is not readily available, such as regions that are conflict-ridden or that experience extreme climate and economic shocks. Doing so would mean relying solely on data that are gathered remotely. The researchers continue to refine and improve the process in order to capture extreme events and areas where access to information is difficult.
Having accurate and timely early warning systems in place is crucial, Michelson notes. “We know that there are irreversible consequences of deprivation to human health and development, especially in early childhood,” she says. That means any improvement that can make intervention faster and more efficient can have large impacts and directly improve the lives of affected populations.
The paper, “A data-driven approach improves food insecurity crisis prediction,” is published in World Development and is available online. [https://doi.org/10.1016/j.worlddev.2019.06.008]
Authors include Erin C. Lentz, University of Texas, and Hope Michelson, Kathy Baylis, and Yujun Zhou, Department of Agricultural and Consumer Economics, College of Agricultural, Consumer and Environmental Sciences, University of Illinois.
Funding was provided by the National Science Foundation, and a USDA National Institute of Food and Agriculture Hatch project, and the University of Texas.