Evaluation of dimensionality reduction techniques in hyperspectral imagery and their application for the classification of terrestrial ecosystems

Abstract

The hyperspectral imagery is formed by a several narrow and continuous bands covering different regions of the electromagnetic spectrum, such as spectral bands of the visible, near infrared and far infrared. Hyperspectral imagery provides extremely higher spectral resolution than high spatial resolution multispectral imagery, improving the detection capability of terrestrial objects. The greatest difficulty found in the hyperspectral processing is the high dimensionality of these data, which brings out the 'Hughes' phenomenon. This phenomenon specifies that the size of training set required for a given classification increases exponentially with the number of spectral bands. Therefore, the dimensionality of the hyperspectral data is an important drawback when applying traditional classification or pattern recognition approaches to this hyperspectral imagery. In our context, the dimensionality reduction is necessary to obtain accurate thematic maps of natural protected areas. Dimensionality reduction can be divided into the feature-selection algorithms and featureextraction algorithms. We focus the study in the feature-extraction algorithms like Principal Component Analysis (PCA), Minimum Noise Fraction (MNF) and Independent Component Analysis (ICA). After a review of the state-of-Art, it has been observed a lack of a comparative study on the techniques used in the hyperspectral imagery dimensionality reduction. In this context, our objective was to perform a comparative study of the traditional techniques of dimensionality reduction (PCA, MNF and ICA) to evaluate their performance in the classification of high spatial resolution imagery of the CASI (Compact Airborne Spectrographic Imager) sensor.