Case study · Agriculture
Weed and crop segmentation across ten growing seasons
Our first client, and still one of our largest: pixel-level outlines of crops and weeds that let a sprayer treat a weed and spare the crop.
- Images segmented since 2016
- 3.8 million
- Crop and weed classes
- 61
- Acceptance, 2025
- 99.1%
The work
The Nairobi centre’s first sixty labellers began by drawing boxes around weeds in sugar beet. Ten years later the work is pixel-level segmentation of 61 crop and weed classes, across images from tractor-mounted cameras in eleven countries, at every stage of growth and in every kind of light.
The hard part is the seedlings. At the two-leaf stage several weeds look almost identical to the crop. We built a guideline with the client’s agronomists in which every confusable pair has its own page of examples, and we run a calibration round at the start of each European growing season.
The client’s sprayers now treat weeds individually rather than spraying a whole field. The client tells us that the labelled data is the single largest factor in the accuracy of their models.