The set of orthogonal eigen-vectors built via principal component analysis (PCA), while very effective for com- pression, can often lead to loss of crucial discriminative information in signals. In this work, we build a new basis set using synthetic aperture radar (SAR) target images via non-negative matrix approximations (NNMAs). Owing to the underlying physics, we expect a non-negative basis and an accompanying non-negative coecient set to be a more accurate generative model for SAR proles than the PCA basis which lacks direct physical interpretation. The NNMA basis vectors while not orthogonal capture discriminative local components of SAR target images. We test the merits of the NNMA basis representation for the problem of automatic target recognition using SAR images with a support vector machine (SVM) classier. Experiments on the benchmark MSTAR database reveal the merits of basis selection techniques that can model imaging physics more closely and can capture inter-class variability, in addition to identifying a trade-off between classication performance and availability of training.