How to set a thematic accuracy target for a national land cover release
Every national mapping agency that publishes a land cover layer eventually gets asked the same question by a reviewer, an auditor, or a minister's office: how accurate is it. Set the target and document it before classification starts. That way the accuracy statement in the release documentation records a decision made in advance, and holds up when a reviewer asks for the confusion matrix.
Setting that target is less about picking a single impressive-sounding figure and more about deciding what the number covers.
Overall accuracy is not the target you need
A single overall accuracy figure (the proportion of all reference points correctly classified) is the easiest number to report and the least useful one for a statistics office. It hides exactly the classes that matter most. A layer can clear a respectable overall accuracy threshold while its built-up class, the one your urban planning department cites in every annual report, sits well below it. Built-up areas are usually a small fraction of total pixels, so a handful of misclassified urban edges barely moves the overall score even as they wreck the one class someone downstream actually relies on.
This is why the convention in land cover work, going back to the original USGS Anderson classification scheme, is to set a per-class target alongside the overall one. Cropland and forest, which dominate most national extents and tend to have distinct spectral signatures, can reasonably be held to a tighter target. Built-up, bare ground, and water's edge, which are smaller, more fragmented, and more likely to be confused with each other at your sensor's resolution, need their own looser target and their own reporting line. Decide the per-class numbers with the department that will cite each class, not in isolation from them.
Producer's accuracy vs user's accuracy: pick the one that matches the question being asked
Once you're past overall accuracy, the next decision is which direction you're measuring in. Producer's accuracy answers one question: of all the reference points that are truly forest, what fraction did the map call forest? It's an omission measure, telling you how much of that class the map is missing. User's accuracy flips the question: of all the points the map labeled forest, what fraction are correct? It's a commission measure, telling someone using the layer how much to trust that label when they see it.
These two numbers can diverge for the same class, and the gap is exactly the information a statistics office needs. Take forest as the example: high producer's accuracy with low user's accuracy means the map is catching most real stands but also mislabeling shrubland as the same class. The reverse pattern is conservative. What gets labeled that way is reliably correct, but patches at the edges go missing. Environmental reporting usually cares more about producer's accuracy, since catching deforestation matters more than a stray false positive, while planning departments pulling the layer for site screening care more about user's accuracy, since they need to trust a built-up label on the parcel in front of them. Set a target for both, per class, and say in the release documentation which one governs which decision.
Writing the target into the accuracy statement for the statutory layer
Once targets exist, the release needs an accuracy statement that a non-specialist reviewer can read: the confusion matrix, overall accuracy, and producer's and user's accuracy per class, against the targets set before production started. Report where a class fell short and why, rather than quietly rounding the overall number up. A statutory layer gets cited by people who weren't in the room when the classification scheme was built, and an honest shortfall documented against a stated target holds up better under audit than a clean-looking number nobody can trace back to a method.
None of this is a one-off exercise. The targets, the reference point design, and the accuracy statement template are work your agency will redo every year the layer gets republished. If your agency is weighing what a single annually-updated, delta-checked classification layer could replace in your current patchwork of classification contracts, Land Use Mapping lays out how that one layer gets produced and handed off to every department that needs to cite it.
Set the targets with the departments who'll use each class, report both accuracy directions, and the next release will be easier to defend than the last one.