Integration of spectral information into support vector machine for land cover classification

Abstract

Support vector machines (SVM) have been widely used for classification purposes. These learning machines are based on classification of data through a kernel function. Classically these kernel functions are either based the Euclidean distance of two data vectors or their dot products. This is a general formulation which is suitable for most data sets. However, when dealing with remote sensing images, the addition of spectral information can add to the divisibility of the data and hence produce higher classification accuracy. In this paper, instead of the Euclidean distance we use the spectral angle function as a differentiation measure of two data vectors. The results show that using this method, high quality separation is achieved leading us to believe that integration of spectral information into the SVM method is indeed an effective approach.