In view of the feature structure of observed data, research is made to improve parametric spectral estimation. Based on linear prediction theory, an optimal method for reflection coefficient estimation is proposed and simultaneously the performance of the algorithm is also analyzed by adopting entropy spectral estimation. The above optimal algorithm is confirmed by computer simulation and has provided a feasible reference for engineering implementation.