This paper proposes to improve classification accuracy of hyperspectral images by using sample interpolation when limited training samples are available. The training data size is artificially increased by adding training samples that have been interpolated from the original training data. Two approaches are presented with different number of training patterns being considered in the interpolation process. In the first approach, the number of samples is approximately doubled, by adding the average of each training sample with another randomly selected training sample of the same class, to the training set. In the second approach, the averages of each sample with each of all other samples of the same class are added to the training set. This approach is referred to as the limit case. For classification, initially, Support Vector Machine (SVM) training is applied to the new and larger sized training data. These support vectors are then used in the classification step. Experimental results show that the proposed algorithm provides increased classification accuracy if a limited number of training samples are available using a simple and effective training data interpolation approach.