Small farms grow much of the world’s food, but from space they are nearly invisible. Their fields are tiny and ill-defined, and the satellite tools built to track crops were designed for the large, uniform fields of industrial agriculture, not the sub-hectare plots that feed many of the world’s poorest people.
New research from the University of Cambridge shows that an AI tool called Tessera could change that. Tessera is a foundation model, trained on years of satellite imagery so that it can be adapted to many different tasks.
Tested on small fields in Austria, Tessera identified most crop types more accurately than methods currently in use while using just 8% of the computing power and none of the hand-tuning those methods require.
Those figures matter because the agencies that plan for food security, among them the UN Food and Agriculture Organization, the World Bank and individual governments, rely on satellite crop maps in their decision making. Surveying small fields on the ground for an entire country is impractical. At that scale a small gain in accuracy can decide whether a country imports enough grain to avoid a shortfall, said lead author Madeline Lisaius, who completed the study as a PhD researcher in Cambridge’s Department of Computer Science and Technology.
“When the decisions are being made at country and continent scale, [that] makes a really big difference in terms of food security and planning,” Lisaius said. “Do we go buy 100 tonnes, or 10,000 tonnes of rice from Thailand now, because we’re going to underproduce and people are going to starve in seven months?”
Lisaius will present the work at ISPRS 2026, a geospatial conference in Toronto on 5 July.
Tessera turns raw satellite images into compact numerical summaries called embeddings, which capture how a patch of ground changes across the seasons. The hardest part of mapping small fields is their edges, where a single satellite pixel can fall across two crops, a track or a hedge, blurring the signal. In fact, according to Lisaius, the smallest fields are almost all edge. That renders them effectively invisible to many methods which rely on looking at a field’s interior to define the crop type, she said.
Tessera embeddings, however, give a signal at the field edge reliable enough to tell one crop from another. The researchers believe this is because the model is trained to track how the land changes across a whole year, rather than reacting to a single, noisy snapshot.
The researchers are careful to state that while the Austrian field trials demonstrate that the technology is effective, using Tessera to guide real decisions where other factors come into play, is still some years off. Nonetheless, people responsible for policy decisions about food security should take the technology seriously now, Lisaius said.
“AI-based technologies like Tessera are very different from anything they’ve ever used before,” she said. “They are cheaper and easier to run than these expensive, complex tools, and with a little investment they could begin to solve some of these really tricky problems within a few years.”
The research was funded by UKRI. A peer-reviewed paper, “Towards Improved Crop Type Classification: a Compact Embedding Approach Suitable for Small Fields,” was authored by Lisaius, Andrew Blake and Srinivasan Keshav (University of Cambridge) with Clement Atzberger (Cyclops.ai) and will appear in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences.
