The bidirectional reflectance distribution function (BRDF) is an important surface property, commonly used to describe reflected light patterns. However, the BRDF is a complex function since it has four angular degrees of freedom and also depends on the wavelength. The direct use of BRDF data set may be inefficient for scene modeling algorithms, for example. Thus, models provide compression and additional functionalities like interpolation. One common way consists in fitting an analytical model to the measurements data set using an optimization technique. But this approach is usually restricted to a specific class of surfaces, to a limited angular or spectral range, and the modeling quality may strongly depend on the optimization algorithm chosen. Moreover, analytical models are unable to actually handle the BRDF dependence on wavelength. We present a new numerical model for acquired spectral BRDF to overcome these drawbacks. This model is based on a separation between the spectral and the geometrical aspects of the BRDF, each of them projected into the appropriate wavelet space. After a brief introduction to the BRDF, the advantages of wavelets and the construction of the model are explained. Then, the performances of modeling are presented and discussed for a large collection of measured and synthetic BRDF data sets. Finally, the robustness of the model is tested with synthetic noisy BRDF data.