The effect of thresholding a maximum likelihood classifier on the accuracy of a landsat classification of a forested wetland

Abstract

Although the maximum likelihood classifier is a popular classification technique, there is an inherent problem associated with the 100% classification of a scene. This is because there will inevitably be pixels within a study area that have a low probability of belonging to any of the predefined categories.

The focus of this research was to locate these low probability pixels and observe their affect on classification accuracy. This was done by performing supervised classifications at various threshold levels using two methods of classification training combined category training site statistics and separated category training site statistics. In general, it was found that a majority of the scene was classified at very low probabilities but the accuracy of the resulting classifications was much greater than the low probabilities would suggest.