Cambridge's open-source Tessera maps Senegal's smallholder crops with 84% accuracy on sparse labels

An 84% hit rate means about 16% of test results were wrong, roughly one in six ^[1]. The phys.org report does not give the crop classes, the rival methods' scores, the basis of the 28% margin, or how the compute comparison treats the cost of generating embeddings ^[4][5].

Tessera splits crop mapping into two steps. In the first, the underlying model reads a year of satellite images and compresses each 10-meter point of land into an embedding, a string of numbers recording how that ground changes over time ^[16]. Then a simple algorithm, calibrated on a few labeled points, turns the embeddings into a crop map ^[17]. Ground labels enter only at the second step, so the method needs few of them. Few labels is what a basin like this one can supply. Plots there are not much larger than a football field ^[14], and current production estimates rest on slow field surveys or expensive processing of satellite imagery ^[15].

Compute splits along the same line. Fitting a simple classifier is a small, repeatable job ^[17]. Producing a year of embeddings for every 10-meter point across a region is a separate job ^[16], and a small agency's bill depends on whether it has to run that job itself.

The cross-year result bears on the constraint Lisaius, the study's lead author, describes. "Accurate and up-to-date crop statistics can guide food security planning and help decide where best to target support. But most local governments and bodies can only afford to collect ground data every few years," Lisaius said ^[7]. "With Tessera, you can train on the data you already have and extend it into the years in between, with more accurate crop information than baseline methods have ever been able to provide." ^[8]

According to the World Food Programme, Senegal's food comes mostly from small, rain-dependent farms, leaving much of the population vulnerable to climate shocks ^[1]. The thing this study doesn't tell you is how a map calibrated in an ordinary season handles an abnormal one. The phys.org report ties the work to this year's El Niño, which it says scientists call the strongest ever recorded, and notes that past strong events brought prolonged drought to West Africa ^[10]. The embeddings record how ground changes over a year ^[16]. If a failed rainy season changes that pattern, a classifier calibrated on a normal year's labels is working outside the conditions it learned from. I'd want to see the transfer tested on a known drought season before gap-year maps are used to target aid.

Lisaius said governments, NGOs and other food security organizations can begin producing their own crop statistics with it now ^[9]. "Reliable agricultural data is essential to anticipate food security and climate-related risks," said Pierre Lucas, the WFP's representative and country director in Senegal ^[11]. His description of the programme's own work is more tentative: "In Senegal, WFP is working with national partners to explore how geospatial data and artificial intelligence can strengthen food security monitoring systems and support faster, more informed decision-making." ^[12]