It is often the case in image classification tasks that image descriptors are of high dimensionality. While adding new, independent, features generally improves performance of a classifier, it increases its cost and complexity. In this paper we investigate how descriptor dimensionality reduction techniques, namely principal component analysis and independent component analysis affect classification accuracy. We test their performance for the task of semantic classification of aerial images. We show that, even with much lower dimensional descriptors, classification accuracy is still near 90%.