The focus of this paper is on automatic annotation for semantic image retrieval. This work is aimed at identifying visual descriptors that are most relevant, effective and suitable for semantic annotation tasks. We propose an image annotation system based on support vector machines and a combination of descriptors that includes a gradient direction histogram and several MPEG-7 visual descriptors. The system is tested on a large database of 7200 cityscape and landscape images. The results indicate that when descriptors are used individually, the proposed gradient direction histogram performs best. However, when descriptors are combined, the accuracy is improved. The presented results confirm that combining the gradient direction histogram and colour structure produces the best results.