How accurately do image classifiers estimate area?

Publisher Website
Google Scholar
Abstract
A spectral image classifier is a rule which relates colour/brightness in a digital image to classes of interest. After an image has been classified, areas can be estimated by counting the number of pixels in each class. This paper presents a simple formula for calculating the accuracy of these area estimates from the confusion matrix. The formula gives the root mean square (r.m.s.) error of the area of class c, in pixel units, as the number or image pixels times the square root of {the sum of the off-diagonal elements in row c and column c of the confusion matrix}, divided by the number of samples in the confusion matrix. The formula is valid for confusion matrices sampled using any non-stratified scheme.

This publication has 5 references indexed in Scilit: