A comparison of canonical discriminant analysis and principal component analysis for spectral transformation.

Abstract

A study was conducted in Michigan's Upper Peninsula to test the strength and weakness of canonical discriminant analysis (CDA) as a spectral transformation technique to separate ground scene classes which have close spectral signatures. Classification accuracies using CDA transformed images were compared to those using principal component analysis [PCA) transformed images. Results showed that Kappa accuracies using CDA images were significantly higher than those derived using PCA at a = 0.05. Comparison of CDA and PCA eigen structure matrices indicated that there is no distinct pattern in terms of source variable contributions and load signs between the canonical discriminant functions and the principal components.