We apply a molecule dictionary approach to synthetic aperture radar canonical feature extraction. These canonical features capture physically-relevant scattering geometry as a function of shape type, frequency, aspect, and polarization. The extraction problem is a nonlinear nonconvex optimization that includes model order selection, feature classification, and parameter estimation. Previous work used image-based initializations, gradient descent, and a hierarchical classification scheme to extract the features. The dictionary approach shifts much of the computational burden to dictionary formation which can be done offline, prior to feature extraction. We show results for cases when the true feature lies in the dictionary and when it does not. Discussion of the practical challenges of dictionary construction is given in the context of recent sparse recovery literature.