The hyperspectral image classification algorithm of relevance vector machine (RVM) is a supervised machine learning algorithm based on Bayes probability model, whose classification accuracy is good and the test time is short. However, the traditional RVM has some shortcomings that the training time will be very long and the effectiveness of the algorithm might decrease if the size of training samples is big or the dimensionality of the data is high. To solve these problems, a hyperspectral image classification algorithm of variational relevance vector machine (VRVM) is proposed. A new distribution is imported into the traditional probability model, which can replace complicated convolution operation with simple logarithm addition operation. Experimental results show that, in the classification of hyperspectral image, the overall classification accuracy and the number of relevance vectors of VRVM are nearly the same with RVM. However, with the increase of the sample, the training time has obviously reduced.