<p>To tackle the obstacles arising from the non-uniformity of coprime arrays and the interference of ocean noise with conventional DOA estimation approaches, an innovative off-grid sparse Bayesian method for estimating the direction of underwater acoustic signals has been devised. This method boosts the precision and resolution in estimating directions through Toeplitz covariance reconstruction and subspace fitting. Firstly, the distribution characteristics of the virtual array elements are analyzed from the perspective of the difference co-array. The covariance matrix is vectorized in conjunction with the corresponding relationship of the path difference obtained by each element thus generating a one-dimensional row vector containing missing data. Subsequently, the maximal consecutive segments of the received signals are intercepted and utilized as input for the Toeplitz reconstruction of the covariance matrix. Next, the signal subspace is solved, facilitating the reconstruction of the received signals. Finally, the Bayesian learning algorithm is invoked to compute the maximal posterior probability associated with the signal. This probability is subsequently employed to ascertain the DOA estimation of the target, facilitating the determination of the target’s orientation. The efficacy and robustness of the proposed methodology are validated through both simulation studies and sea trial experiments.</p>