A map of farmland leaves an obvious question unanswered: what is growing there? For anyone trying to understand a region’s food supply, the difference between a staple crop, a cash crop and a field left fallow matters. A green patch on a satellite image supplies only part of the answer.
Research publicised by Cambridge on 29 September applies its Tessera model to that problem in Senegal’s groundnut basin. The work, led by Madeline Lisaius, examines whether a representation learned from satellite observations can make local crop mapping easier to repeat.[4]
The field changes; the representation follows
These are small, varied plots. More than 92% of the labelled fields in each study year were below two hectares. The 2018 reference data included 551 cropland polygons; 2019 had 486. Their categories covered groundnut, millet, sorghum, cowpea and fallow land. A polygon records the outline of a surveyed field, connecting an image to something identified on the ground.[1][2]
Tessera compresses a pixel’s changing satellite observations into 128 numbers at roughly 10-metre resolution. Its learning procedure encourages two different samples from the same cloud-free time series to produce matching representations. The result is a reusable description of the temporal signal, incorporating radar and optical observations. Local researchers can download these precomputed descriptions instead of processing every observation afresh.[3]
The point of watching a year is that two crops can look similar on one date and develop differently over a season. A useful representation can retain information about that changing appearance even when some observations are missing. It still needs a local connection between the pattern and the crop’s name.
The Senegal study supplies that connection with supervised classifiers trained on field labels. All pixels belonging to one polygon remain in the same training, validation or test partition.[1]
Learning one year, recognising another
For same-year crop classification, Tessera’s mean accuracy was 84.6% in 2018 and 69.4% in 2019. It reached 56.7% in the different, eight-class 2021 task. Those are distinct label sets and tasks, not a controlled decline over time.[1]
Crop-classification accuracy when the training year changes
| Training → test year | Tessera | AlphaEarth | Raw time series |
|---|---|---|---|
| 2018 → 2019 | 62.6% | 36.4% | 16.1% |
| 2019 → 2018 | 55.5% | 45.0% | 21.0% |
The transfer result is the practical attraction. A field can change crop, and a growing season can shift, without requiring the mapping team to begin its entire modelling effort again. The model receives new satellite observations; it reuses what it learned about relating those observations to crop labels. The remaining errors also show why transfer cannot simply be assumed.
The reference labels set the level of detail the map can promise. Intercropping was recorded inconsistently between years, so the analysis uses the primary crop. It cannot inventory every species growing together in a plot. The supplement also finds that errors cluster within fields: a hundred neighbouring pixels are not a hundred independent confirmations.[2]
A shared computation with a local purpose
In the separate land-cover computation test, Tessera used about 78% less CPU time than the raw-data approach; AlphaEarth used less again. This compares downstream classification using ready-made inputs, excluding the earlier work of generating the embeddings.[1]
That earlier work is substantial. Tessera’s documentation describes generating embeddings as a demanding processing job, while recommending downloads of existing coverage for downstream applications. Its software is released under MIT terms, and the embeddings and model weights under CC0. Availability varies by region and year, so the reusable layer is useful where the required observations have already been processed.[3]
For a regional mapping team, this changes where effort can go. Repeatedly assembling satellite features is work that can be shared; identifying what farmers actually planted remains local. Combining the two could make a crop map easier to maintain through successive seasons, giving small fields a more regular place in the picture of a farming region.
Sources & further reading
- Towards Accessible Smallholder Crop Classification in the Groundnut Basin of Senegal — accepted manuscript Environmental Research: Food Systems; University of Cambridge repository · Accessed 17 Oct 2026
- Supplementary materials: Towards Accessible Smallholder Crop Classification in the Groundnut Basin of Senegal University of Cambridge repository · Accessed 17 Oct 2026
- TESSERA — project documentation University of Cambridge / GitHub · Accessed 17 Oct 2026
- How new AI tool could help efforts to improve food security in the Global South University of Cambridge, Department of Computer Science and Technology · 29 Sep 2026 · Accessed 17 Oct 2026