How many components are needed to represent the spectral reflectance of a surface? What is the dimension of a spectral reflectance? How many image channels are needed for the acquisition of a multispectral colour image? Such and similar questions have been discussed extensively in the literature. We have done a survey of the literature concerning this topic, and have seen that there is a large variation in the answers. We propose a method to quantify the effective dimension of a set of spectral reflectances. The method is based on a Principal Component Analysis, and in particular on specific requirements for the accumulated energy of the principal components. We apply the analysis to five different databases of spectral reflectances, and conclude that they have very different statistical properties. The effective dimension of a set of Munsell colour spectra is found to be 18, that of a set of natural object reflectances 23, while the effective dimension of a set of reflectances of pigments used in oil painting was only 13.