Due to its intuitive nature, the decoding of covert visuospatial attention has recently been proposed to be useful for Brain–Computer Interface (BCI) applications. In order to identify neural correlates of covert spatial visual attention, state of the art approaches usually rely on the whole α–band over fixed time intervals. In this work, we propose a discriminative model that exploits spectro-temporal evolution of covert visuospatial attention to improve classification performances. Results with 10 healthy subjects demonstrate that our approach reaches, on average, 0.74±0.03 of AUC value with an increase (+11.5%) with respect to the state of the art method. In addition, the proposed method allows faster classification (<1 second on average) without compromising classification performances.