Abstract: Seismic external source coherent noise (ESCN) is generated by vibrations other than the main seismic source, and it overlaps significantly with effective signals in both the time domain and the frequency domain. ESCN tends to obscure the effective signals, thereby interfering with the effectiveness of seismic exploration. To suppress ESCN, this study proposes an unsupervised learning method based on similarity measurement. The study uses the average cosine similarity function to determine the apparent velocity of the noise, obtain the noise propagation direction, and calculate the proportion of the noise in the original data, thereby deriving the predicted ESCN. There are differences in amplitudes between the predicted ESCN and the true ESCN. Unsupervised deep neural networks, with excellent nonlinear mapping capability, can correct amplitudes of the predicted ESCN and obtain the amplitude-corrected estimated value of ESCN as well as the results of effective signals by minimizing the mean absolute error loss function. The proposed unsupervised learning method, which does not rely on true labeled data, can effectively address the issue of missing training datasets in field data acquisition, and exhibits a broad applicability. Examples of synthetic and field data demonstrate that the proposed method in this study can effectively suppress ESCN, with its performance superior to that of the traditional similarity measurement method, the conventional frequency-wavenumber (FK) filter method, and the Karhunen-Loeve (KL) filter method.