Cortical activity can be estimated from electroencephalogram (EEG) or magnetoencephalogram (MEG) data by solving an ill-conditioned inverse problem that is regularized using neuroanatomical, computational, and dynamic constraints. Recent methods have incorporated spatio-temporal dynamics into the inverse problem framework. In this approach, spatio-temporal interactions between neighboring sources enforce a form of spatial smoothing that enhances source localization quality. However, spatial smoothing could also occur by way of correlations within the state noise process that drives the underlying dynamic model. Estimating the spatial covariance structure of this state noise is challenging, particularly in EEG and MEG data where the number of underlying sources is far greater than the number of sensors. However, the EEG/MEG data are sparse compared to the large number of sources, and thus sparse constraints could be used to simplify the form of the state noise spatial covariance. In this work, we introduce an empirically tailored basis to represent the spatial covariance structure within the state noise processes of a cortical dynamic model for EEG source localization. We augment the method presented in Lamus, et al. (2011) to allow for sparsity enforcing priors on the covariance parameters. Simulation studies as well as analysis of real data reveal significant gains in the source localization performance over existing algorithms.