This paper presents a new criterion to deal with training data heterogeneity in space-time adaptive processing (STAP). It is based on Siegel measure, which is a distance in the space of Hermitian positive definite matrices, and consists in computing the distance between the covariance matrix estimated from training data and an a priori matrix. After deriving a test based on this distance, the statistical behavior of Siegel distance is analyzed, as a function of the number of training samples. Its probability density function (PDF) is derived and related to that of the generalized inner product (GIP) in the case of one single training snapshot. Finally, simulations are performed to illustrate the interest of preprocessing training data using Siegel distance, in terms of STAP signal to interference plus noise ratio (SINR) performance