This study investigates the effect of training set selection strategy on classification accuracy of hyperspectral images. This effect is analyzed in conjunction with three other factors, namely the use principal component analysis on the input data, and the use of spatial information and choice of classifier. Support Vector Machines (SVM) and Maximum Likelihood (ML) classifiers are used for demonstration. Meanshift segmentation and majority voting are used for inclusion of spatial information. The effect of the training data size and sampling strategy is demonstrated over the high resolution Pavia University hyperspectral data.