• FEWS NET scientists are using machine learning to improve crop yield forecasting in data-sparse regions, allowing analysts to predict agricultural production even in countries with limited or inconsistent historical data.
• By leveraging relationships between weather, soil conditions, and crop performance across multiple countries, researchers can transfer insights to similar regions, significantly improving forecast accuracy where local data are lacking.
• This innovative approach, supported by new datasets and research, is helping close critical agricultural data gaps and strengthening early warning systems to better anticipate food insecurity and guide humanitarian response.

While scientists with the Famine Early Warning Systems Network (FEWS NET) can't predict the future, they can accurately forecast crop yields. Their powerful models analyze a range of factors, from weather patterns to soil conditions, to predict how well an agricultural season will perform, even before planting begins.

Crop yield forecasts are critical components of FEWS NET’s scenario development process that analysts use to warn of food emergencies months before they occur. 

“FEWS NET would generally like to know what volume of grains will be produced and harvested in the countries we monitor,” UCSB Climate Hazards Center (CHC) researcher Frank Davenport said. “For a lot of populations we study, that has a direct relationship to their food security.”

But across Sub-Saharan Africa, there is currently an agricultural data gap that sometimes prevents FEWS NET partner organizations like the CHC from getting the data they need to feed crop yield forecasting models.

“This is a general problem that’s been known for a while,” Davenport said. “The data is just so spread out and inconsistent.”

Recently, FEWS NET convened a Crop Production Working Group that aims to tackle this problem and standardize crop production data for wide, public use. In 2024, the group cleaned and uploaded a total of 4.7 million individual harvest results, comprising 18.8 million data points for 37 countries.

But there are still some places where data scarcity is a concern. These places are often remote, underdeveloped, and sometimes stricken by conflict. 

“Most of the data we’re using is collected by countries themselves as part of their annual agricultural census,” Davenport said. “But in places like Somalia, for example, that have experienced periods of instability, we have data for some years but not others.” 

An enormous amount of data – from both the past and present – are needed to forecast crop yields. These data include historical observations of crop yields, ground-based weather station recordings of rainfall and soil moisture, satellite-based observations of things that can impact crops like rainfall, temperature, evaporation, and more.

Without sufficient data, the accuracy of crop yield forecasts can decline. 

To solve this issue, CHC scientists developed an innovative machine learning approach that takes existing data from one country and applies them to another that has geographically and climatically similar features.

By leveraging machine learning – a subset of statistical modeling – researchers can better observe the nature of relationships between two phenomena and apply that understanding elsewhere. For example, scientists may look at how high rainfall levels may relate to high crop yields, or vice versa.

“Machine learning really takes that type of relationship and scales it up a lot, both computationally and algorithmically, so we can take lots of variables and explore all the different combinations they might have with each other and all the different ways they can influence yields,” Davenport said. “We might ask, ‘How do hot and wet days, or dry and cool days, or any combination therein come together to influence crop yields?’”

Shraddhanand Shukla, a researcher with the CHC, further explained how this new machine learning approach allows relationships between indicators to be tested without any prior assumptions.

Davenport, Shukla, and Donghoon Lee, a former CHC postdoctoral researcher and now an assistant professor at the University of Manitoba, recently published two papers that evaluate the efficacy of FEWS NET’s new machine learning approach to crop yield forecasting.

In the first paper, researchers used Earth Observation (EO) data, such as rainfall and vegetation indicators, combined with maize yield data from Burkina Faso, Kenya, Malawi, and Somalia to train machine learning models and test their transferability across regions.

They found that models performed better in areas with medium to high rainfall that were trained on data from multiple countries. 

“What this suggests to us is there is a high potential for being able to train models on a set of countries, and then being able to predict grain yields in the medium to higher producing areas of countries where we don’t have data,” Davenport said.

In the second paper, researchers used EO data and machine learning to evaluate how different types of data and modeling frameworks influence forecast accuracy for maize yields in Burkina Faso and Somalia.

Lee explained that by combining data on unchanging factors like soil type and livelihood zone with variable data on changing weather patterns, they could better understand how climate and weather affect different farming regions. This approach also improved the accuracy of their corn yield predictions.

Looking toward the future, Davenport, Shukla, and Lee emphasized that this progress, alongside ongoing in-season crop yield forecasting practices, will provide critical insights to address food security concerns in FEWS NET-monitored countries and help to mitigate the impacts of climate variability. 

”Yield forecasting in FEWS NET-monitored countries will always be a challenge, but we are at a real pivot point where the data, machine learning methods, and computational infrastructure have all converged and opened up new avenues for research and application that simply were not there even 5 years ago,” Davenport said.

Sign up to receive FEWS NET reports and updates in your inbox.