Abstract: Random noise is one of the common background noises in seismic data, and its attenuation will directly affect the signal-to-noise ratio of seismic data, which is of great significance to improve the quality of seismic data. Low-rank approximation technique is a commonly used method to suppress random noise of seismic data. It converts frequency spatial domain data into the form of Hanke matrix, and uses singular value decomposition technique to reconstruct data by retaining large singular values, so as to achieve the purpose of rank reduction and suppress random noise. The method takes advantage of the low-rank nature of noiseless seismic data, which can be destroyed in the presence of random noise. However, traditional singular value decomposition technology has low computational efficiency, and seismic data generally consist of a large amount of datasets, so traditional singular value decomposition technology will inevitably lead to a large increase in time cost. In order to improve the efficiency of random noise suppression, a new singular value decomposition technique based on compressed sensing theory is proposed. The sparse representation of data is considered in the calculation of singular values, and the sparse representation of data is used to approximate the solution of high-dimensional singular vectors and singular values, so as to improve the accuracy and computational efficiency of singular value decomposition. Compressed sensing theory makes full use of data sparsity, avoids direct processing of original high-dimensional data, and theoretically has high computational efficiency. Three-dimensional synthetic seismic records and field data examples are used to verify the validity and practicability of the proposed method, and comparisons with traditional and random singular value decomposition techniques are performed. The results show that the improved low-rank approximation technique can effectively suppress random noise in seismic data, and the effective signal can be enhanced and highlighted. Compared with traditional and random singular value decomposition, the compressed singular value decomposition technique has higher computational efficiency and can greatly save time cost. Low-rank approximation technology based on compressed singular value decomposition has better performance than other methods in random noise suppression and can further improve the signal-to-noise ratio.