Seismic Random Noise Attenuation via Implicit Neural Representations With Multiple Structural Priors

Abstract

Seismic data are inevitably contaminated by random noise arising from various sources, which seriously degrades the quality of subsequent data processing, inversion, and interpretation. Deep learning has emerged as an effective approach for seismic random noise attenuation. However, these methods are often susceptible to overfitting, which limits their denoising performance. In this letter, we integrate three structural priors into a unified framework to alleviate overfitting and enhance denoising capability. Specifically, low-rank matrix factorization parameterized by implicit neural representations (INRs) is employed to reconstruct 2-D seismic data, simultaneously embedding low-rank and smoothness priors. Two additional INRs are further introduced to learn the Dirichlet energy (DE), which measures column- and row-wise similarities in seismic data and imposes a nonlocal similarity prior. Extensive experiments on 2-D synthetic and field seismic datasets demonstrate that the proposed method consistently outperforms existing deep learning-based approaches.