A Deep Ground Roll Noise Suppression Method Based on DRinCGAN

The deep seismic reflection data obtained from actual field exploration contains rich geological information, but also suffers from severe interference, especially low-frequency strong ground roll. This kind of noise is widely distributed and intertwined with reflected waves, seriously damaging the characteristics of effective reflection signals and leading to a decrease in signal-to-noise ratio, which affects the quality of the data. In order to improve the signal-to-noise ratio and support subsequent seismic exploration research, it is necessary to suppress ground roll. In recent years, deep learning technology has provided a new approach for ground roll suppression. However, creating high-quality target seismic data training sets is time-consuming and traditional methods have limitations. To solve these problems, a ground roll suppression method based on DRinCGAN is proposed, which combines disentangled networks and recurrent generative networks, designs feature encoders to extract effective reflection waves and noise features, and designs a ground roll suppression generator to obtain processed seismic records. Compared with supervised learning methods, this method does not require paired data and is more suitable for real seismic records. Theoretical analysis, numerical simulation, and practical processing results show that the method is effective and feasible, enriches the theory of pre-stack ground roll suppression, and expands the technical means for processing complex seismic data.