Spectral Angle Based Kernels for the Classification of Hyperspectral Images Using Support Vector Machines

Abstract

Support vector machines (SVM) have been extensively used for classification purposes in a broad range of applications. These learning machines base their classification on the Euclidean distance of the data vectors or their dot products. These measures do not account for the spectral signature information that can be achieved from remote sensing images. Given the high value of this information, integrating it into the SVM algorithm is a reasonable suggestion. This paper utilizes the spectral angle (SA) function as a measure for classification of a hyperspectral image. The SA function is joined together with the radial basis function (RBF) to form a spectral angle based RBF function. Experimentation results are promising and confirm that this approach can compete with existing classification methods.