Devising efficient sparse decomposition algorithms in large redundant dictionaries has attracted much attention recently. However, choosing the right dictionary for a given data set remains an issue. An interesting approach is to learn the best dictionary from the data itself. The purpose of this contribution is to describe a new dictionary learning algorithm for multichannel data analysis purposes under specific assumptions. We assume a large number of contiguous channels as in so-called hyperspectral data. In this case it makes sense to consider a priori that the collected data exhibits sparse spectral signatures and sparse spatial morphologies in specified dictionaries of spectral and spatial waveforms. Building on GMCA, the proposed algorithm gives a practical way to enforce the additional a priori spectral sparsity constraint on the dictionary space. Numerical experiments with synthetic and real hyperspectral data illustrate the efficiency of the proposed algorithm.