It has been recognized that the performance of the traditional direction of arrival (DOA) estimation algorithms is highly dependent on array geometry parameters, especially the array size and element number. Unfortunately, significant performance degradation may be caused under unideal array geometry such as small size array onboard small underwater vehicle. Among all the possible directions, the presence of underwater target at a certain direction constitutes a type of spatial sparsity. Thus, spatial sparse recovery provides a potential way to address this challenging problem. As each DOA will correspond to a different transmission response pattern with respect to each array element, thus constitutes a direction-dependent transmission response (DDTR) that can be explored. However, random contribution of underwater acoustic (UWA) channel contained in DDTR makes it extremely difficult to directly construct dictionary matrix with DDTR. In this paper, a novel DDTR based small size array DOA estimation method is proposed. First, DDTRs are decomposed into time delay (TD) term and random channel response (RCR) term to decouple the adverse impact of UWA channel. Second, TD is conveniently constructed as dictionary matrix in terms of geometry relationship. Then, by accumulating the constructed DOA at frequency bins selected with energy or correlation coefficient, a selective frequency accumulation compressed sensing reconstruction (SFA-CS) algorithm is designed to mitigate the negative effect of UWA channel. Compared with the conventional DOA estimation algorithms, simulation results demonstrate that the proposed method can achieve small array DOA estimation with a lower root mean squared error (RMSE) and higher success rate (SR).