The spatial correlativity and spectral information are not applied synchronously in the classification model of hyperspectral data. To solve this problem, an improved classification approach based on Markov random field (MRF) theory is proposed in our work. The MRF model based on maximum a posteriori is applied to combine the spectral and spatial information. The probabilistic support vector machine (PSVM) algorithm using pixels spectral information is applied to improve the estimation accuracy of the class conditional probability (CCP) of the spectral energy function, and the efficient belief propagation (EBP) based on multi-accelerated strategy (such as ordinal propagated message strategy, linearized message-updating strategy, and coarse-to-fine approach) is developed in order to solve the problem of the high calculational complexity and time-consumed in the global energy minimum optimization of MRF model. The true hyperspectral data collected by airborne visible infrared imaging spectrometer (AVIRIS) is applied to estimate the performance of the proposed approach in the agricultural demonstration area, Indiana northwest, USA. The performance of the proposed approach is compared with simulated annealing and iterated conditional model. The results illuminate that the average classification accuracy of our method reachs to 95.78%, and the Kappa coefficient is 93.34%, much higher than that of the result by the traditional MRF classification algorithms, and the computational efficiency is improved more than 3 times compared with the belief propagation algorithm.