A new approach to AI crop mapping has shown that satellite imagery can identify smallholder fields in Senegal with useful accuracy even when researchers have relatively little labelled local data. That matters because some of the places where better crop maps could help most are also the places where field-by-field training datasets are hardest and most expensive to build.
The work focuses on Senegal’s Groundnut Basin, where small fields, mixed cropping and year-to-year changes make agricultural mapping difficult. Instead of training a fresh model from scratch on one season of local labels, the researchers used Tessera, a foundation model designed to turn earth-observation imagery into reusable numerical representations of the landscape.
AI crop mapping normally needs more local labels than farmers can provide
Satellite images contain far more information than a human mapmaker can inspect field by field. The difficulty is teaching a model what each patch means. A supervised crop classifier usually needs examples where the true crop is already known, often from field surveys or carefully checked records.
That creates a structural problem for smallholder agriculture. Fields can be tiny and irregular. A single plot may contain more than one crop. Boundaries can shift. Cloud, planting dates and seasonal conditions can change what the same crop looks like from space. A model that works well in one year may weaken when moved to another.
Tessera tries to reduce the cost of that local learning step. Its project materials describe a pixel-wise earth-observation foundation model: a system pre-trained to create general representations from satellite data that can then be adapted to a specific task with less labelled data than a specialist model would normally require.
This is similar in spirit to the way large language models learn broad patterns before being adapted to a narrower job. The inputs are very different, but the practical goal is familiar: do expensive general learning once, then reuse it across many local problems.
Senegal gives the model a difficult real-world test
The researchers evaluated crop classification in the Groundnut Basin using Sentinel satellite imagery and ground observations. Their comparison looked not only at which method achieved the highest accuracy, but also at whether it could work with limited labels, transfer across years, run with realistic computing resources and remain accessible enough for practical use.
The study found that methods using Tessera representations performed strongly across those criteria. In one temporal-transfer comparison reported in the paper, the Tessera approach was 28% more accurate than the next-best method. That comparison is specific to the tested years and evaluation setup, so it should not be read as a universal performance margin for crop mapping.
Temporal transfer is a particularly useful stress test because agricultural agencies need maps that survive a change of season, not models that work only on the year used to train them.
The cross-year test is particularly important. Agricultural monitoring is useful only if it can be repeated. A model that has to be rebuilt from a large new ground survey every season may be too expensive for routine use. Reusable representations can make annual updating more practical, even when performance still changes with the quality of the new season’s labels.
LiveAIWire has previously covered the evidence behind AI in precision agriculture. The Senegal work fits a slightly different problem. It is not trying to tell one farmer exactly how much fertiliser to apply. It is trying to make the landscape itself more legible at scale.
A crop map can become public infrastructure
Knowing what is growing where has uses far beyond producing a colourful map. Governments and development organisations can use crop information to estimate production, understand drought exposure, plan food-security responses and target field surveys more intelligently. Researchers can combine crop maps with rainfall, soil and market data to study how agriculture changes over time.
In regions dominated by small farms, conventional agricultural statistics can be slow or expensive to collect. Satellite mapping does not replace ground surveys, because the model still needs reliable observations for training and validation. It can, however, help those limited field measurements cover a much larger area.
That makes label efficiency a practical issue, not simply a machine-learning benchmark. Every ground-truth point can mean somebody travelled to a field, identified a crop correctly and recorded it in a usable form. A method that can learn more from fewer trustworthy observations changes the economics of the whole mapping exercise.
The approach also sits beside other uses of orbital data. LiveAIWire reported on AI systems detecting methane plumes from space. Both stories show a broader shift in earth observation: satellites provide the raw view, while machine learning increasingly determines which patterns can be extracted quickly enough to guide action.
The limitations are exactly where future deployments will struggle
The Senegal study does not show that one model can be dropped into any country and immediately identify every crop. The researchers found that performance can fall when moving between years, in part because the quality and distribution of ground-truth data change. The work also did not fully solve the problem of secondary crops in multi-crop fields.
Those caveats matter because smallholder farming is diverse by definition. A field boundary that looks clear in one region may be meaningless in another. Crop calendars differ. Intercropping can make a single label misleading. Satellite resolution can also limit what can be distinguished when plots are extremely small.
Foundation models reduce the amount of local information required; they do not abolish the need for local knowledge. A national agricultural service would still need people who understand the crops, seasons and survey quality well enough to decide whether the map is trustworthy.
There is also a policy question around openness. Tessera is presented as an open system, which could make it easier for research groups and public agencies to adapt the technology without paying for a closed commercial platform. That does not make deployment free, because imagery processing, training, verification and staff still cost money, but it lowers one barrier to experimentation.
Better satellite AI is making small places visible
A striking feature of modern earth-observation AI is that the frontier is moving from broad categories toward fine-grained local detail. Earlier satellite analysis often worked best on large features such as forests, cities or major crop belts. Foundation models are designed to capture reusable patterns at much smaller scales.
That same capability can support environmental monitoring, disaster response and land-use planning. LiveAIWire has looked at AI wildfire detection using satellites, another case where the value comes from converting a continuous stream of remote-sensing data into a timely map of something people care about on the ground.
The Senegal result is encouraging because it tests the idea where data scarcity is not an academic inconvenience but part of the real problem. A crop-mapping system for smallholder agriculture has to cope with fewer labels, smaller fields and changing seasons or it will remain a laboratory exercise.
Tessera does not remove those constraints. It makes them more manageable. If that holds across more countries and crop systems, the most important effect may be simple: agricultural maps that were once too expensive to update regularly could become routine enough to support decisions every season.
There is a second reason this type of mapping matters: agricultural policy often needs information before a harvest is complete. A late national statistic can describe what happened, while a reliable seasonal map can help agencies decide where to inspect, where a drought may be affecting a particular crop and which districts deserve closer attention. The value comes from updating the picture while decisions can still change.
That does not mean a satellite model can see yield directly or diagnose every problem on a farm. Crop identity, crop condition and final production are different measurements. A map can be a foundation for those analyses, but it still has to be combined with weather, field observations and local agronomy. Keeping those boundaries clear prevents an accurate classification system from being sold as a complete agricultural forecasting tool.
If future tests show the same approach transferring across more seasons and countries, the payoff could be cumulative. Each new field survey would not start from zero. It could refine a reusable mapping system, gradually improving the local evidence base while keeping the cost of annual updates within reach.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.