Vector sensor imaging presents a challenging problem in covariance estimation when allowing arbitrarily polarized sources. We propose a Stokes parameter representation of the source covariance matrix which is both qualitatively and computationally convenient. Using this formulation, we adapt the proximal gradient and expectation maximization (EM) algorithms and apply them in multiple variants to the maximum likelihood and least squares problems. We also show how EM can be cast as gradient descent on the Riemannian manifold of positive definite matrices, enabling a new accelerated EM algorithm. Finally, we demonstrate the benefits of the proximal gradient approach through comparison of convergence results from simulated data.