This paper proposed a new subspace based method to estimate the direction of arrival (DOA). the existence of coherentsources results in the rank deficit of sample covariance matrix. Classic MUSIC(Multiple Signal Classification) can not classify coherentsources, and instead, generate an equivalent sources somewhere between them. In our method, at first, the contribution of the equiva-lent sources is subtracted by the collected data. And then, the method is recursively applied. Simulation experiment show that com-pared to the other method, this method has higher spatial resolution and less estimation error. This paper proposed a new subspace based method to estimate the direction of arrival (DOA). The existence of coherents results in the rank deficit of sample covariance matrix. Classic MUSIC (Multiple Signal Classification) can not classify coherents, and instead, generate an equivalent In our method, at first, the contribution of the equivalence to lent sources is subtracted by the collected data. And then, the method is recursively applied. Simulation experiment show that com-pared to the other method, this method has higher spatial resolution and less estimation error.