In order to make full use of the abundant spectral information and spatial information of hyperspectral images, this paper proposes a hyperspectral image classification algorithm based on dominant sets clustering and Markov random fields. First of all, the local spectral-spatial consistency of hyperspectral images is analyzed, the measurement of both band informativeness and independence is completed, an un-directed weighting graph is constructed and dominant sets clustering method is used to select the optimal band subset which preserves good structure information. Secondly, the local spectral-spatial consistency of adjacent pixels after the band selection is established by using Markov random fields, which makes the context information of the image space effectively used. Finally, according to the Bayesian theorem, the hyperspectral image classification problem is transformed into the maximum posterior probability which can solve the problem and yield the classification results. Experiments on two datasets, i.e., Indian Pines and Pavia University, show that this algorithm can achieve higher overall classification accuracy and Kappa coefficient than other similar algorithms.