Based on the time-frequency distributions, a novel subspace method is proposed in this paper to estimate the directions-of-arrival (DOA) of non-stationary signals. The array time-frequency correlation matrix C/sub x/, instead of the traditional correlation matrix R/sub x/, is constructed and eigen-decomposed to separate the signal subspace and the noise subspace. By processing the observed data in both spatial domain and time-frequency domain simultaneously, the "signal selectivity" is achieved. Therefore DOA of signals with identical frequency band and time slot but with different time-frequency signatures can be extracted one by one, and the estimation performances are significantly improved over the conventional method. It is also proven herein that the traditional subspace methods are just the low-dimensional special cases of the proposed approach.