Hyperspectral imaging (HSI) is used in analysis of paintings to obtain features hidden to the human eye by selecting specific wavelengths. Superpixel segmentation can be applied to HSI for feature extraction. A superpixel algorithm processes an image in a way in which the result includes an unnecessary amount of over-segmentation. In this work, we use over-segmentation and propose Spectral Similarity Merging (SSM), a region growing algorithm based on homogeneous spectral properties with the aim to reduce over-segmentation without compromising under-segmentation. The algorithm focuses on the similarity of the spectral shapes rather than intensity. Results show an average of 45% reduction in oversegmentation and an average of 53% improvement on the Fscore on existing superpixel segmentation algorithms.