Satellite crop maps guide subsidies and food aid, so agencies check them against crops recorded on the ground. Those checks are rarely fair: field teams label fields beside roads and skip small or remote ones, so the score describes only the easy places.
The fix is prediction-powered inference: the map's guesses across the whole region give a first estimate, and its mistakes on the labelled fields correct that estimate, weighted so fields from poorly covered areas count for more.
In five European regions where every field is registered, biased checks overstated the accuracy score (macro-F1, a per-class average) by up to 0.20; correcting cut the worst error to 0.07.