In brain-computer interfaces, adapting a classifier from one user to another is challenging but essential to reduce training time for new users. Common Spatial Patterns (CSP) is a widely used method for learning spatial filters for user specific feature extraction but the performance is degraded when applied to a different user. This paper proposes a novel Adaptive Selective Common Spatial Pattern (ASCSP) method to update the covariance matrix using the most probable candidates. Subspace alignment is then applied to the extracted features before classification. The proposed method outperforms the standard CSP and adaptive CSP algorithms proposed by other authors. Visualization of extracted features is provided to demonstrate how subspace alignment contributes to reduce the domain variance between source and target domain.