30 September 2026, Cambridge: A satellite-based artificial intelligence tool developed at the University of Cambridge has correctly identified crop types across Senegal’s groundnut-growing region with 84 percent accuracy, using far less ground-truth data than competing approaches typically require, according to a study published September 29 in the journal Environmental Research: Food Systems.
The tool, called Tessera, was developed by a team in Cambridge’s Department of Computer Science and Technology led by Madeline Lisaius, working with co-authors S. Keshav, A. Blake and C. Atzberger. It was tested in partnership with the World Food Programme in Senegal’s groundnut basin, a region where smallholder farmers grow groundnut alongside millet, sorghum and other staples on small, irregularly shaped plots that are notoriously difficult for satellite-based crop mapping systems to interpret accurately.
How the System Works
Most AI systems for classifying crops from satellite imagery need large volumes of labeled training data, meaning fields where researchers already know exactly what crop is growing, in order to learn to recognize patterns. That kind of detailed ground-truth data is scarce and expensive to collect across the smallholder farming systems that dominate much of Africa and South Asia, which is one reason satellite crop monitoring has been slower to reach these regions than large, mechanized farms in North America or Europe.
Tessera takes a different approach. It is what researchers call a foundation model, a type of AI system trained first on a very large amount of general satellite data to learn broad patterns, which can then be adapted to specific tasks with much less additional data. The Cambridge team’s version analyzes a full year of satellite imagery for each 10-meter patch of land and converts the changing patterns it sees, how the land’s reflectance shifts through planting, growth and harvest, into a compact numerical summary called an embedding. Once that conversion is done, only simple, lightly calibrated algorithms are needed to sort those embeddings into crop types, sharply reducing the amount of labeled ground data and computing power required.
Testing the system against imagery from 2018, 2019 and 2021, the researchers found it correctly classified crop types 84 percent of the time and outperformed the next-best comparison model by 28 percent in one scenario, while using a fraction of the computing resources. Notably, the model also performed well when trained on data from one year and then applied to a different year it had never seen, a capability that matters in smallholder regions where field boundaries, planting patterns and even which crop is grown can shift from season to season. The research was funded by UK Research and Innovation and Mantle Labs, alongside the World Food Programme collaboration.
Global Agriculture is an independent international media platform covering agri-business, policy, technology, and sustainability. For editorial collaborations, thought leadership, and strategic communications, write to pr@global-agriculture.com
