Use of the average mutual information index in evaluating classification error and consistency

Abstract An information theory measure of shared information, average mutual information (AMI), is applied to the comparison of thematic maps. The use of a posteriori entropies for one map, given the class identity from the second map, allows evaluation of individual class performance. Examples are presented using classification error matrices and overlays of maps with different themes. AMI measures a different aspect of the problem than does either Percentage Correct or Kappa, measuring ‘consistency’ rather than ‘correctness’. Because of their different viewpoints in comparing classifications, and the potential for spotting mislabelling problems, one should use a combination of Kappa and AMI to assess error matrices. AMI provides a means of assessing the similarity of maps with different themes.