Seismic Random Noise Suppression with Wavelet-Inspired Invertible Network

Abstract

Random noise attenuation is an essential process in seismic data processing, with  a major impact on subsequent workflows such as seismic interpretation. Because of the variety of frequency bands and apparent velocities, suppressing seismic random noise is challenging. Conventional techniques for suppressing seismic random noise include vector-median filtering, low-rank approximation and sparsity-domain based techniques. Degradation of effective seismic signals is inevitable, especially in areas with extremely complex structures. Machine learning is widely used in seismic data processing, offering high efficiency and automation. However, the black-box nature of machine learning algorithms limits their application in seismic data processing. To address this problem, we propose applying the multi-scale Wavelet-inspired Invertible Network (WIN) to seismic random noise suppression. The network has the following advantages: (1) It combines multiple lifting invertible neural networks, which ensures perfect reconstruction of the original signal; (2) The multiple lifting invertible neural networks act as multiple scales of the original signal and resemble denoising in the wavelet domain in a more efficient way; (3) The denoising performance of the proposed network can be controlled by choosing the noise level. A numerical study on field seismic data shows the advantage of the neural network over conventional denoising algorithms.