Support vector classification of remote sensing images using improved spectral Kernels

Abstract

A very important task in pattern recognition is the incorporation of prior information into the learning algorithm. In SUppOlt vector machines this task is performed via the kernel function. Thus for each application if the right kernel function is chosen, the amount of prior information fed into the machine is increased and thus the machine will perform with much more functionality. In the case of hyper-spectral imagery the amount of information available prior to classification is a vast amount. Current available kernels do not take full advantage of the amount of information available in these images. This paper focuses on deriving a set of kernels specific to these imagery. These kernels make use of the spectral signature available in images. Subsequently we use mixtures of these kernels to derive new and more efficient kernels for classification. Results show that these kernels do in fact improve classification accuracy and use the prior information available in imagery to a better degree.