Evaluation of Factors Affecting Support Vector Machines for Hyperspectral Classification

Remote sensing data are attractive for deriving land cover information through image classification. A number of parametric and non-parametric classifiers such as the maximum likelihood classifier (MLC) and the artificial neural network (ANN) have been developed and tested successfully on multispectral data. However, the existing classifiers have shown marked limitations in the classification of hyperspectral images obtained from sensors such as AVIRIS, HYMAP, HYDICE and MODIS. Recently, Support Vector Machine (SVM) has been proposed as an alternative for classification of both multi and hyperspectral data. SVM is a machine learning algorithm that employs an optimizer to identify an optimal separating hyperplane to discriminate two classes of interest. The results from a few studies on the use of SVM for remote sensing image classification are promising and encouraging. However, there are several issues that need to be considered and investigated before SVM becomes operational in remote sensing applications. This paper presents some results from an ongoing study on SVM based classification of remote sensing data. The aim is to investigate the effect of some factors on the accuracy of SVM classification. The factors considered are selection of multiclass method, choice of the optimizer and the type of kernel function. The results show that among different multiclass methods, optimizers and kernel functions, classification performed with Directed Acyclic Graph multiclass method using the RBF kernel function produced the highest accuracy of 97% when Lagrangian SVM is used as the optimizer.