Seismic random noise attenuation with deep skip autoencoder based on hybrid attention mechanism

Abstract

Denoising seismic data is a crucial step in seismic data processing to enhance the signal-to-noise ratio of data because random noise is inevitably introduced during seismic data acquisition owing to environmental factors. In this study, we introduce a symmetric skip-connected denoising method (A-SK22) based on a hybrid attention mechanism with a hybrid pool to attenuate noise in seismic data. The proposed method adopts the coding-decoding network structure of the U-Net network. In the encoding phase, hybrid pooling is employed to reconstruct seismic data more effectively, mitigating the risk of partial loss of valid information during down-sampling. The network structure of hybrid pooling consists of a parallel arrangement of the average and maximum pooling. In the skip-link part, the sum operation, which reduces the computational cost, is adopted. Meanwhile, in pursuit of further mining the spatial and channel information of the seismic data, we added the global attention mechanism in the skip linking part. The recovery experiments conducted with synthetic and actual seismic data demonstrate the effectiveness of the proposed method in attenuating random noise while causing minimal distortion to essential seismic signals.