The purpose of this paper is to increase the efficiency of spectral analysis (i.e. to reduce the root mean square error (RMSE) of direction-of-arrival (DOA) estimation based on root similarity approach initially proposed by A. Gershman. The used methods are: spectral analysis methods, pattern recognition methods, digital statistical modeling methods. The following results were obtained. The root classification approach is used in the case of joint application of two types of data covariance matrix (standard covariance matrix (CM) and estimate of CM with Toeplitz structure). This approach removes the outliers (outlying roots) from preliminary DOA estimates (roots corresponding to the preliminary DOAs). The modification of initial root classification approach is proposed. It consists of avoiding averaging of DOA estimates obtained by estimator for the different CM at high signal-to-noise ratios (SNRs). This step for considered case allows to improve the performance of DOA estimation using the root classification approach. Simulation results are presented confirming the performance of proposed approach. Conclusions. The performance improvement of the subspace-based methods of spectral analysis can be attained by removing the outliers from the initial DOA estimates. The simultaneous application of classical second-order CM and estimate of structured CM gives two sets of DOA estimates (roots of polynomials). Root classification approach processes these sets and improves the performance of DOA estimation. The modification proposed in the paper gives the additional advantage at high SNR. The considered approach is also can be used together with other polynomial rooting methods of DOA estimation. K e ywor d s : Direction-of-arrival estimation; Karhunen-Loève transformation; spectral decomposition of correlation matrix; spectral analysis methods; pattern recognition.