Prior-Driven Self-Supervised Lightweight Method for Seismic Signal Denoising

TL;DR

The paper tackles seismic noise attenuation by introducing a lightweight self-supervised adaptive convolutional filter (ACF) with 2464 learnable parameters. It leverages two priors—a local prior to suppress high-frequency noise and a global variance prior to limit signal leakage—along with a low-scale learning strategy to enhance performance without requiring large labeled datasets. The approach is validated on both 2D/3D synthetic data and field data, where ACF demonstrates competitive or superior noise suppression while preserving important seismic features, and it maintains computational efficiency and interpretability. The authors provide open-source code to facilitate reproducibility and practical deployment in seismic exploration.

Abstract

Seismic exploration is currently the most mature approach for studying subsurface structures, yet the presence of noise greatly restricts its imaging accuracy. Previous methods still face significant challenges: traditional computational methods are often computationally complex and their effectiveness is hard to guarantee; deep learning methods rely heavily on datasets, and the complexity of network training makes them difficult to apply in practical field scenarios. In this paper, we proposed a neural network that has only 2464 learnable parameters, which is hundreds or even thousands of times lower than that of the current mainstream deep learning networks. And its parameter constraints rely on priors rather than requiring training data. We proposed two types of priors: the local prior and the global variance prior for self-supervised learning, and put forward low-scale learning to further enhance its performance in noise processing. We validated our method on both synthetic and field data, and the results indicate that our proposed approach effectively attenuates random noise.

Figures (5)

• Figure 1: a) The workflow of our proposed method; b) the structure of ACF; c) the downsampling and merging methods of downsampling learning.
• Figure 2: 3D synthetic seismic data. a) and b) represent the clean seismic data and the noisy seismic data, respectively; c) to f) represent the processed results of ACF, TDAE, DRR, and HSLR; g) to j) represent the absolute errors between clean data and results of ACF, TDAE, DRR, and HSLR.
• Figure 3: SNR of four methods' results on synthetic data.
• Figure 4: Noisy field seismic data. a) represents the noisy field seismic data; b) represents a local enlarged view of the area within the red box; c) to f) represent the processed results of ACF, TDAE, DRR, and HSLR; g) to j) represent the local enlarged view of results of ACF, TDAE, DRR, and HSLR.
• Figure 5: RMS and ALS of four methods' results on field data.