Antenor Blog 3 min read
You Can Count Crops From Space — Accurately Enough to Trust
Open satellite imagery plus machine learning turns 'how much land is planted' into a measured number. Validated against a government's own sown-area figures, our classifier reached over 90% accuracy — no proprietary data required.
Governments, insurers and agribusinesses spend heavily to answer a simple question: how much land is actually planted, and with what. Satellites already photograph every field every few days, and machine learning can turn that raw imagery into a hectare count you can trust. Ours reached over 90% accuracy measured against the sown-area figures a national agriculture secretariat reports — built entirely on open Sentinel-2 data, no proprietary sensors.
“How much of it is there?” is one of the most expensive questions in physical operations. How many hectares are planted. How much of the parcel is built. How much of the flood plain is under water. The usual answers come from field surveys and self-reported declarations — slow, costly, and inconsistent with each other. Meanwhile a satellite has already imaged the whole area, for free, on a schedule.
The gap is not data. It is turning multispectral imagery into a number that holds up against ground truth.
Why this is a machine-learning problem, not a photo
From orbit, a crop is not obvious. It looks different by region, season and growth stage; clouds, bare soil and water all masquerade as signal; and “green” is not the same as “planted.” A human labeling pixels does not scale to a province, and a fixed threshold breaks the moment conditions change.
A trained classifier handles exactly this: it learns the spectral and temporal signatures that distinguish crops from everything else, across regions, and produces a per-area estimate instead of a pretty map. The same pipeline extends to other land classes — water bodies, flood zones — by adding labeled examples.
The part that makes it trustworthy: external validation
Most “AI accuracy” numbers are measured against a held-out slice of the model’s own data. That tells you the model is consistent with itself, not that it is right about the world.
The number that matters is validation against an independent, authoritative benchmark. In our crop-classification work, the classifier’s planted-area estimates were checked not against a test split but against the sown hectares a national agriculture secretariat officially reports — an external source with no connection to the model. Reaching over 90% agreement there is a fundamentally stronger claim than a high score on your own held-out data. When you evaluate a model, the question is always: right compared to what?
Why “open data only” is a feature, not a limitation
The whole system runs on Sentinel-2 L2A imagery — free, public, global. No proprietary satellite contract, no exclusive sensor feed. That matters for two reasons: the cost floor is low enough to make routine monitoring viable, and the approach is portable — the same method that measures cropland can be pointed at land use, construction progress, deforestation or disaster mapping.
Capability that used to require a data budget most operations do not have is now a modeling problem on data anyone can access.
Where this applies beyond agriculture
- Insurance and lending: verify planted area or land use behind a policy or loan without a site visit.
- Supply chain and commodities: independent read on regional production, ahead of official reports.
- Environment and government: monitor deforestation, flooding or land-use change at scale.
- Infrastructure: track construction and site changes over time from imagery.
We build these as predictive and computer-vision models on open remote-sensing data, validated against real-world ground truth — because an estimate you cannot check is not a measurement.
Frequently asked questions
How accurate is satellite crop classification, really?
It depends on the crop, region and validation method, but the meaningful figure is agreement with an independent ground truth, not a model’s own test split. Our classifier reached over 90% against official sown-area statistics — a benchmark the model never trained on.
Do you need paid or proprietary satellite data?
No. Open Sentinel-2 imagery is free, global and updated every few days, and it is enough for crop, water and land-use classification. Paid sub-meter imagery helps for tasks that need fine spatial detail, but most area-measurement problems do not.
Can the same approach map things other than crops?
Yes. The pipeline is a classifier over multispectral imagery; changing the labels changes the target. The same method maps water bodies, flood zones, land use, construction and vegetation change, which is why remote sensing generalizes across agriculture, insurance, environment and infrastructure.