Cognitive radio networks achieve spectrum sharing by utilizing the idle periods of licensed bands via dynamic spectrum access technique. Spectrum characterization and prediction help perform more efficient spectrum sensing and then optimize spectrum access strategy. In the paper, UTD-HDP, a nonparametric Bayesian model, was introduced by extending the standard HDP(Hierarchical Dirichlet Process) to perform utilization data clustering and distribution parameters estimation. Using this model, we characterized the features of spectrum utilization adaptively and predicted the future spectrum utilization with high accuracy.