This paper proposes a novel algorithm to estimate the time-varying spectral centroid of the narrowband Doppler volume backscattering signal. It is constructed in a semi-parametric way, that is, modeling parametrically the local narrowband evolutionary spectrum using an AR(2) and nonparametrically adapting its time-varying coefficients using the wavelet shrinkage. The improved performance is gained by the underlying linear prediction function of the AR(2) and the minimax optimality of the unknown smoothness adaptation of the wavelet shrinkage procedure.