Application of transfer learning and multi-scale feature fusion in intelligent suppression of seismic random noise
Xin Xu, Wuyang Yang, Wei Xinjian, Haishan Li, Nang Wang · 2024
Denoising Convolutional Neural Networks (DnCNN), a data-driven learning algorithm, has been widely applied in suppressing Gaussian noise and enhancing super-resolution reconstruction in recent years. This paper focuses on optimizing the DnCNN framework for seismic noise suppression, aiming to cater to the intelligent noise suppression needs of extensively acquired threedimensional (3D) seismic data. We propose an advanced 3D seismic denoising framework based on DnCNN, which extends the network architecture from two dimensions to three dimensions and incorporates concepts of transfer learning and multi-scale feature fusion. Addressing the challenge of not being able to fully capture real noise in practical data, we employ conventional processing for denoising 3D post-stack data, creating labels and datasets with original seismic data and the noise obtained from processing. DnCNN utilizes a residual learning strategy, taking noisy seismic data as input and predicting seismic random noise as output. Subtracting the latter from the former yields the denoised actual seismic data post-stack. The novel method proposed in this paper has been tested and applied in various working areas, demonstrating its ability to suppress random interference and other types of noise, enhance the continuity and signal energy of the target layer reflections, effectively prevent denoising artifacts, and significantly improve the efficiency of poststack denoising technology.