Radiation source localization is very important in electromagnetic environmental monitoring activities. Most current positioning methods can only obtain source position estimates at the center of fixed grids, and these methods usually require prior knowledge of the monitoring environment. In this paper, a new variational Bayesian-based radiation source positioning method is proposed, which dose not need the prior knowledge of the environment. We modify dictionary parameters at multiple stages, and thus, the estimated locations can be at any positions in the grids. In addition, we select the self-information of different observation data and eliminate relevant interferences. As a result, the computational complexity can be reduced while guaranteeing the localization accuracy. Simulation results show that the proposed method can improve the localization accuracy effectively compared with the state of the art.