Abstract: A dictionary learning-based method for sound speed profile (SSP) inversion in shallow water is presented. The empirical Green's functions between two parallel horizontal arrays can be extracted from noise cross-correlation functions. The sound speed profiles are sparsely characterized by data-generated dictionary matrix, and they can be inverted by searching for sparse coefficients. This method is validated by experimental data in the South China Sea. Compared with the traditional empirical orthogonal function (EOF) methods, the inversion accuracy accuracy is reduced to 0.53 m·s^-1, moreover this method has fewer search parameters and higher accuracy.