Abstract: The time-frequency analysis is combined with array processing to develop a DOA (Direction of Arrival) estimation method. The array data model is constructed in time-frequency domain by cross time-frequency distribution of the reference sensor and other sensors of the two symmetric sub-arrays. Accordingly the subspace approach is performed on the averaging of the two sub-arrays' time-frequency data vector instead of the traditional array model, to estimate the DOAs of multiple signals. Because the array data is processed both in spatial domain and 2-D time-frequency domain, the proposed method has the ability to select the signal of interesting, and is suitable for non-stationary signals. Additionally, the method is robust to noise and holds the advantage of low computational burden. Simulations are conducted to verify the efficiency of the method and comparision is made with other methods.