The increasing interest of the construction aggregates industry in reducing production costs and the costs resulting from improper use of construction materials leads to the question whether it is possible to statistically identify some rock variants by their reflectivity of near-infrared and mid-infrared light. Infrared spectroscopy allows quantitative and qualitative analysis of minerals in a reliable manner, whereas the classification of rocks is complicated by the fact that the optic behavior of minerals forming the rock often appears muted. In addition, minor constituents may dominate the spectrum. Furthermore the relevant spectra form high dimensional data, which are extremely difficult to analyse statistically, especially when curves are very similar. Common methods of multivariate statistics for this type of data, used in chemometric studies, followed by linear discriminant analysis, do not lead to acceptable classification error rates. In this paper wavelets are used in order to reduce dimensionality. As wavelets are better able to mirror local behavior of curves, they are more suitable for selecting characteristic features. The approximation is analyzed in terms of its classification properties using Mahalanobis distance or flexible discriminant analysis.