This paper proposes a novel knowledge-aided approach for selecting training data in space-time adaptive processing (STAP) whose performance suffers from a severe degradation in heterogeneous interference environment. The proposed approach exploits distances between interference covariance matrices of training data and tested data as the measurements of interfer- ence statistical similarities, which helps us gain a deeper insight into the statistics from the point of geometry. Three distances including Euclidean distance, Riemannian distance and a physical distance are combined to distinguish various heterogeneous phenomenons. A prior knowledge is employed in estimating the interference covariance matrices of both training data and tested data. Simulation results illustrate the effectiveness of the proposed approach.