Multi-class extensions of the GLDB feature extraction algorithm for spectral data

The generalized local discriminant bases (GLDB) algorithm proposed by Kumar, Ghosh and Crawford in (2001), is an effective feature extraction method for spectral data. It identifies groups of adjacent spectral wavelengths and for each group finds a Fisher projection maximizing the separability between classes. The authors defined GLDB as a two-class feature extractor and proposed a Bayesian pairwise classifier (BPC) building all pairwise extractors and classifiers followed by a classifier combining scheme. With a growing number of classes the BPC classifier quickly becomes computationally prohibitive solution. We propose two alternative multi-class extensions of GLDB algorithm, and study their respective performances and execution complexities on two real-world datasets. We show how to preserve high classification performance while mitigating the computational requirements of the GLDB-based spectral classifiers.