The applicability of the existing spectral estimation methods is restricted to the stationarity conditions in a wave field. For nonstationary conditions a generalization of the conventional definition of spectrum as a distribution in both time and frequency suggests a filtering technique for spectral estimations. The statistical estimation errors associated with the general form of estimates are concisely described in a relative mean square error criterion. A minimization of this error as an overall figure of merit for the accuracy of estimates yields an optimal procedure for constructing an estimate that can be regarded as the most accurate one for the given physical wave data. Concepts are illustrated with actual wave data analysis, and the validity of the optimal procedure is demonstrated by simulation and comparison with the results from conventional techniques.