Nonlinear feature extractor for unsupervised classification of hyperspectral image

Hyperspectral image has high spectral dimension, vast data and altitudinal interband redundancy. It causes problems to image classification. In order to effectively reduce dimensionality and improve classification precision, we presented a new unsupervised classification approach, which exploited curvilinear distance analysis (CDA) method as nonlinear feature extractor and employed independent component analysis mixture model (ICAMM) as unsupervised classifiers. We applied this approach for unsupervised classification of a test image from the Airborne Visible/Infrared Imaging Spectrometer and used classification accuracy to evaluate the quality of classification. The experiments demonstrate that the classification accuracy resulted from this approach is higher than that from conventional techniques, indicating the efficiency of CDA in feature extraction of hyperspectral imagery.