In order to further improve the DOA estimation accuracy of coherent signals, a modified algorithm based on main feature vectors is proposed. First, the main eigenvectors corresponding to the signal subspace are selected, and the inverse conjugate transformation is performed to obtain the augmented main feature eigenvectors. Then, the linear prediction equation is constructed, and the weighted least squares method is used to solve the polynomial coefficients in the prediction equation. Finally, the DOA of the signals is obtained by finding the root of the polynomial. The modified algorithm overcomes the weakness of PUMA algorithm under conditions of the deteriorated severely performance and the completely cohered signals, and is good in performance. It’s no matter whether the signals are completely coherent or partially coherent. The maximum number of the resolvable signals are improved. The high precision can be obtained when the signaltonoise ratio is low and the number of snapshots is small. Compared with the PUMA algorithm, when the signal is partially coherent, the root mean square error of the proposed algorithm is reduced by about 0.2°; When the signal is completely coherent, the root mean square error is reduced by about 0.8°.The performance of the algorithm is verified by comparison with various decoherence algorithms.