To solve the problem of performance degradation due to nonhomogeneous clutter in adaptive radar processing, a new training data selector in non-homogeneous compound-Gaussian clutter based on polarization knowledge was proposed. The polarization scatting matrix of every training sample was estimated using maximum likelihood estimation (MLE) method, and then the error between the estimation and prior polarization knowledge was used to remove outliers from training data. The performance of the knowledge-based algorithm was analyzed on simulated radar data. The results show that the new data selector removes outliers effectively when outlier-clutter ratio is low and achieves a satisfactory performance level for the estimation of clutter covariance matrix.