Cortical source imaging aims at identifying activated cortical areas on the surface of the cortex from the raw EEG data. This problem is ill-posed, the number of channels being very low compared to the number of possible source positions. In some realistic physiological situations, the active areas are sparse in space and of short time durations, and the amount of spatio-temporal data to carry the inversion is then limited. In this work, we propose an original data driven space-time-frequency dictionary which takes into account simultaneously both spatial and time-frequency sparseness while preserving smoothness in the time-frequency (i.e., non-stationary smooth time courses in sparse locations). Based moreover on these assumptions, we take benefit of the Matching Pursuit (MP) framework for selecting the most relevant atoms in this highly redundant dictionary. We apply two recent MP algorithms, Single Best Replacement (SBR) and Source Deflated Matching Pursuit (SDMP), and we compare the results using a spatial dictionary and the proposed Spatial-Time-Frequency (STF) dictionary to demonstrate the improvements of our multidimensional approach. We also provide comparison using well established inversion methods, FOCUSS and RAP-MUSIC, analysing performances under different degrees of non-stationarity and signal to noise ratio.