The ability of remote sensing data to acquire measurements of land surfaces at different spatial

spectral and temporal scales renders it a major source of land cover information.The process of relating pixels in a satellite image to known land cover is referred to as image classification.Support Vector Machines（SVMs） represent one of the new generation land cover classification techniques.There are three commonly used SVMs namely:linear

polynomial and radial basis function（Gaussian） SVM classifiers

whose successful deployment is dependant on the selection of respective optimum parameters.The voluminous nature of remote sensing data renders the identification of these parameters slow and tedious.In this paper

a new data reduction technique that optimizes remote sensing data for SVM classification is proposed.This research shows that the quantity of data needed to derive the SVM parameters can be reduced without adversely affecting the land cover classification accuracy.Data reduction was successfully applied to polynomial and radial basis function（RBF） SVM classification.Linear SVM classification results however

resulted in significantly poorer classification accuracies.