Abstract: Because the direction of arrival (DOA) estimation mathematical model of the grid-less sparse reconstruction method is in the form of single snapshot, the method has superior performance only when the noise level approaches zero. In order to improve the performance of the grid-less method when the signal to noise ratio (SNR) is low, a multi-snapshots grid-less DOA estimation method is proposed. First, we perform singular value decomposition (SVD) to get the time-domain signal subspace of the observation matrix and achieve noise reduction of the observation matrix by projection from the observation matrix to the time-domain signal subspace. Then, in order not to increase the computational complexity of multi-snapshot, the column vectors of the observation matrix are weighted and accumulated to obtain the single snapshot form. Finally, it is proved theoretically that the model solved by the GL-SVD method proposed in this paper is convex and can achieve the precise reconstruction of wideband signal DOA. The sim- ulation results show that the proposed method has higher estimation accuracy than L₁ norm minimum singular value decomposition (L₁-SVD) and off-grid sparse Bayesian inference singular value decomposition (OGSBI-SVD) with low SNR and wideband coherent sources. In addition, it has higher estimated probability and resolution in the case of smaller angle interval.