Interpreting the change detection error matrix

Abstract

Two different matrices are commonly reported in assessment of change detection accuracy: (1) single date error matrices and (2) binary change/no change error matrices. The third, less common form of reporting, is the transition error matrix. This paper discuses the relation between these matrices. First, it is shown that the transition error matrix implicitly measures temporal correlation in classification errors. Based on two assumptions (no correlation, maximum correlation), the single date error matrices can be used to obtain a most pessimistic and most optimistic estimate of the transition accuracy. Next, it is shown that the change/no change error matrix does not quantify certain classification errors. It is shown that change/no change error matrix can be used complementary to the full transition error matrix in efforts to improve transition detection accuracy. Despite its advantages, the transition error matrix is only very rarely reported, while it is of interest to all those interested in the accuracy of transitions (from–to) in change detection.