National crop maps in Canada and the United States are produced by classifying satellite images with machine learning. Agriculture and Agri-Food Canada (AAFC) publishes an Annual Crop Inventory at 30-meter resolution with a design target of at least 85 percent overall accuracy, and the USDA National Agricultural Statistics Service (NASS) publishes a Cropland Data Layer whose measured overall crop accuracy was 75.4 percent for 2025. The two figures describe different things, so they should not be compared directly.
Key takeaways
- AAFC's inventory began in the Prairie Provinces in 2009 and became a national product from the 2011 growing season.
- NASS reports overall crop accuracy of 81.6 percent for 2023, 77.5 percent for 2024, and 75.4 percent for 2025, alongside producer's accuracy of 85 to 95 percent for major crops.
- NASS says counting pixels in its layer is biased downward and is not a substitute for official acreage estimates.
- Satellite maps describe land cover. They do not report yield, soil health, or what a farmer intended to grow.
What a satellite crop map is
A crop map is a grid in which every cell carries a label such as canola, spring wheat, corn, or pasture. The labels are not read from photographs by eye. A classifier learns how each crop changes in color and texture across a growing season, using images taken at several dates, and then assigns a label to every cell. Fields that look alike in July can differ clearly in May or September, which is why imagery is collected throughout the season rather than once.
The learning step needs examples where the true crop is known. In the United States, NASS takes agricultural training and validation data from the Farm Service Agency's Common Land Unit program. Cells that are not farmland borrow their labels from the USGS National Land Cover Database. A model is therefore only as good as the reference data behind it, and the non-farm labels on the map are only as accurate as the database they come from.
How the Canadian and US products compare
The table lists what each agency states on its own pages. The accuracy entries are not the same kind of number, and they cover different crops and regions.
| Item | AAFC Annual Crop Inventory | USDA Cropland Data Layer |
|---|---|---|
| Cell size | 30 meters (0.09 hectare); 56 meters in 2009 and 2010 | 10 meters (0.01 hectare) from 2024; 30 meters for 2008 to 2023 |
| Classifier | Decision tree methodology using optical and radar imagery | Decision tree (See5) for 2008 to 2023; random forest in Google Earth Engine for 2025 |
| Accuracy statement | Overall target accuracy of at least 85 percent | Measured overall crop accuracy: 81.6 percent (2023), 77.5 percent (2024), 75.4 percent (2025) |
| Update | Annual, with years listed from 2009 through 2025 | Annual; the 2025 layer was released February 27, 2026 |
AAFC's record describes 85 percent as a target of the method. The record does not give a measured value, and a reader who needs one should look for the agency's accuracy assessment for a specific year and province.
Reading the accuracy numbers
Overall accuracy is the share of checked cells whose map label matches the reference label. Producer's accuracy asks a narrower question: of the cells that truly grow a given crop, what share did the map label correctly. NASS states that producer's accuracy is generally 85 to 95 percent for the major crop categories, so a lower overall figure is not a contradiction. Minor crops and non-farm classes can pull the overall number down.
NASS also notes that its earlier accuracy assessments used buffered validation data, which inflated the results somewhat, and that unbuffered assessments began with the 2016 layer. The 2025 layer switched from a decision tree to a random forest classifier. Year-to-year changes in the published accuracy therefore mix changes in the map with changes in method and in how accuracy is checked.
Where the maps are used and where they mislead
NASS describes the Cropland Data Layer as a way to provide supplemental acreage estimates to its Agricultural Statistics Board. The limits are practical. Adding up mapped cells gives a raw number that NASS says needs bias correction, and counting pixels is usually biased downward compared with official figures. Small or narrow fields are also harder to label at 30 meters than large uniform ones, because edge cells mix the crop with roads or hedgerows. An illustrative example: a 3-hectare field covers about 33 cells at 30 meters, and a large share of them sit on its border.
A checklist for using a crop map
- Find the year, cell size, and classifier named in the product documentation.
- Check whether the stated accuracy is a target or a measurement.
- Look at accuracy for the specific crop and region, not only the overall figure.
- Confirm whether the validation data were buffered or independent.
- Use official statistics for totals and the map for spatial patterns.
FAQ
Does the crop map show how well a crop is growing?
No. Both products label which crop is on the ground. Yield and crop condition are separate estimates that rely on other data, and the maps do not report them.
How often do the maps update?
Both are annual. AAFC's inventory is listed for 2009 through 2025 on the federal open data portal, and the NASS 2025 layer was released on February 27, 2026, after the growing season.
Can the map reveal what a specific neighbor grows?
It is a cell-level summary with known errors, so it is not a reliable record of any single field. Both agencies document accuracy for the product as a whole, not for individual parcels.
Is the data free to use?
AAFC lists the federal open government license for its inventory. NASS provides its layer as is and says it does not warrant results, so users are responsible for their own interpretation.
Related reading
- Where artificial intelligence is applied in agriculture
- Sustainability effects of AI on environment and society
- Traditional methods compared with modern machine learning
- How machine learning training is carried out
More in AI and Society.
Last verified: October 2026