Abstract: A novel wavelet packets-based DOA (direction of arrival) estimation method is proposed for the rapid performance degradation of typical multiple signal classification (MUSIC) algorithm in the scenario where different signals' DOA are close together even coherent, i.e., the subband-based MUSIC algorithm (SB-MUSIC). The SB-MUSIC decomposes the fullband signal into subbands and chooses the optimal leaf nodes in accordance to the best basis criterion. Then, the MUSIC method is applied to every leaf node chosen so as to estimate individually the subband spectrum. The advantages of subband spectra, such as improving SNR and amplifying frequency spacing, are analyzed taking account of that when using SB-MUSIC only the fullband makes sense, thus giving a way for frequency mapping from subband back to fullband. Simulation results showed that SB-MUSIC will not only improve the resolution of classical eigenstructure-based methods but also decouple the coherent signals if the fullband is properly decomposed into subbands. In particular, the rate of success of estimation and spectral flatness measure (SFM) of subband are obviously higher than that of fullband.