Multispectral sensor design using performance measure-based hyperspectral band grouping

This paper introduces the concept of using a performance measure to select band groups when using a single classifier. Typically, band groups are selected using a proximity measure to determine the similarity or dissimilarity of hyperspectral bands. The problem with that approach is the similarity or dissimilarity of the hyperspectral bands may not be important for some problems. The novelty of using a performance measure is it can be used to select band groups that result in the best performance regardless of the similarity between the band groups. The results demonstrate better overall accuracy using our approach when compared to uniform partitioning. The improved performance is realized when using a Support Vector Machine or Bayes Maximum Likelihood classifier. Often these improvements are achieved using fewer band groups than uniform partitioning.