USF Policy Lab Projects

Improving Farm and Fishery Forecasts Using Machine Learning

Faculty: Jesse Antigua-Hughes, David Saah, Fernanda Lopez Ornelas, Andrew Hobbs

Food security agencies need to know early when harvests are going to fall short. Right now that information comes from famers surveys and ministry crop statistics that are slow, expensive, and often innaccurate. We're building a system that predicts crop yields directly from satellite imagery so humanitarian agencies and governments can spot trouble months before harvest. 

Our team trains machine learning models on satellite data (Landsat, Sentinel, MODIS), weather, records, soil maps, and harvest statistics from 35 African countries (GROW-Africa), covering roughly 3,250 districts over the last 25 years. The result is a public web map at hobbservations.com/africa-yields that shows maize yield estimates and forecasts up to six months ahead, with calibrated uncertainty bounds for every prediction. 

The policy stakes are high. Recent cuts to USAID funding have weakened the data infrastructure humanitarian agencies rely on for early warning, just as climate variability is making harvests less predictable. Cheap, automated yield forecasts can help fill that gap, supporting agencies and partner governments as they target food assistance, position commodity reserves, and price drought insurance before crisis arrive.