The Watson distribution defined on the hypersphere is one of the most used distributions for modelling axial data. In this paper, we consider the discriminant analysis for axial data assumed to come from a mixture of Watson distributions defined on the hypersphere. We develop an optimal classification rule, which enables us to assign a new observation into one of several Watson subpopulations defined on the hypersphere. As the probabilities of misclassification cannot be calculated in closed form, we report on a simulation study to estimate, in some cases, the probabilities of misclassification and a distance between the two Watson subpopulations defined on the hypersphere. An illustration of our approach is provided using data defined on the sphere given in the literature.