This paper investigates the parameter estimation of compound Wishart distribution for statistical modeling of polarimetric synthetic aperture radar data. We show that the recently proposed method of matrix log-cumulant may lead to non-invertible equations and cause unstable performances. In order to overcome such difficulty, we proposed a Bayesian-based method to re-estimate the log-cumulants. Simulation experiment demonstrates that the proposed algorithm provides both improved and stabler results. Finally, we present an application of the new method for texture analysis of the Germany F-SAR polarimetric data.