Remote Sensing
Confusion matrix
A table cross-tabulating mapped classes against reference classes, the basis for accuracy statistics.
The matrix shows, for each mapped class, how many reference samples were correctly and incorrectly assigned. From it come overall accuracy, user's accuracy and producer's accuracy, each answering a different question.
Publishing the matrix alongside the map is what makes an accuracy claim checkable. A single overall figure hides whether errors concentrate in the class that matters most for the carbon estimate.
Frequently asked questions
What is a good overall accuracy?
It depends on the classes and the decision at stake. A high figure can still conceal poor performance in a rare class.
Why report both user's and producer's accuracy?
One describes the reliability of a mapped class, the other how much of the real class was captured.