123
BEYOND ACCURACY ASSESSMENT:
CORRECTION OF MISCLASSIFICATION
Nicholas R. Chrisman
University of Wisconsin
Madison, WI 53706
ABSTRACT
Map accuracy analysis is critical to remote sensing, but
currently accepted techniques, such as percentage correctly
classified, provide ambiguous information and rest on shaky
statistical foundations. Cohen's kappa, a measure of
agreement, provides a more comparable index of
classification success. Map accuracy data should be applied
to more than accuracy assessment; it provides the basis for
correcting area estimates. This paper demonstrates
Tenenbein's double sampling method for correcting errors in
classification, a simple manipulation of conditional
probabilities. Results from published data show substantial
changes in area estimates.
BACKGROUND
The assessment of accuracy has become a common topic in
studies of remote sensing. Unfortunately, this work seems
divided into two distinct populations. One set grapples the
statistical questions of designing samples, but seems
removed from actual data problems. The other set is firmly
based in application projects, but not particularly
conversant with sophisticated analysis. Information
generated in applied accuracy studies is not fully utilized.
Furthermore, research in remote sensing seems isolated from
similar work in statistics. There is a need to pull these
strands together to ensure better tools for application
projects. This paper will attempt to reevaluate the methods
available and to contribute some further possibilities.
METHODOLOGICAL WORK IN ACCURACY ANALYSIS
The question of overall map accuracy was addressed in a
fairly comprehensive manner by Hord and Brooner (1976).
They decomposed the overall problem into errors at control
points, along boundary lines, and for classification of
areas. This general approach is useful beyond remote
sensing circumstances, but concern in this field has been
focussed almost exclusively on the area case. Either point
and line errors are considered smaller than the
classification errors, or, perhaps, the point and line
errors are bundled into a technology which must be primarily
judged by its classification accuracy.
Hord and Brooner emphasize the probabilistic aspect of
classification error and suggest reporting the percentage
correct with a statistically parametric confidence interval.
They recognize the binomial distribution for true-false
data, but regard the normal approximation as adequate for