An unsupervised random noise suppression method in frequency domain for 3D seismic data

Improving the signal-to-noise ratio is a key step in seismic data processing. The current deep learning- based noise reduction methods have achieved better results. However，these methods are carried out in the temporal- spatial domain based on the local similarity of the seismic data and the processing efficiency is low. In view of the lateral continuity of geological structure，the shot gathers are very similar. Thus，an unsupervised rank-reduction denoise method in frequency domain is proposed based on the low-rank feature of the same frequency component of 3D data. The low-rank principle in frequency domain of 3D data is expounded and the singular value decomposition theory is used to guide the establishment of autoencoding network；Considering the characteristics of random noise distribution in frequency domain，K-L（Kullback-Leibler）divergence is used to constrain the loss function to improve the denoising effect. The experiments on synthetic and field data verified the advantages of the proposed method in denoising performance and computational efficiency compared with the multichannel singular spectrum analysis （MSSA）and K-SVD（K-Singular Value Decomposition）methods. © 2023 Editorial office of Oil Geophysical Prospecting. All rights reserved.