Distance to second cluster as a measure of classification confidence

Abstract

Most image classification algorithms rely on computing the distance between the unique spectral signature of a given pixel and a set of possible clusters within an n-dimensional feature space that represents discrete land cover categories. Each scrutinized pixel will ultimately be closest to one of the predefined clusters; different classification algorithms differ in the details of which cluster is considered as closest or most likely, but in general the selected algorithm will label each pixel with the label of the closest cluster. However, pixels expressing virtually identical distances to two or more clusters identify a limitation of this typical classification approach. Conditions for limitations to distance based classification algorithms include when distances are long and the pixel may not clearly belong to any single category, may represent mixed land cover, or can be easily confused spectrally between two or more categories. We propose that retention of the distance to the second closest cluster can shed light on the confidence with which label assignment proceeds and present several examples of how such additional information might enhance accuracy assessments and improve classification confidence. The method was developed with simplicity as a goal, assuming the classification has already been performed, and standard clustering reports are available. Over a test site in central British Columbia, Canada, we illustrate the described technique using classified image data from a nation-wide land cover mapping project. Calculation of multi-spectral Euclidean distances to cluster centroids, standardized by cluster variance, allows comparison of all potential class assignments within a unified framework. The variable distances provide a measure of relative confidence in the actual classification at the level of individual pixels.