Deploying Self-Supervised Learning for Real Seismic Data Denoising

2026-05-11

AI Summary

This study addresses the challenge of denoising real seismic data, where supervised methods are hindered by the absence of clean reference signals. Building upon the Noisy-as-Clean (NaC) self-supervised learning framework, the work abandons conventional synthetic Gaussian noise and instead injects controllable real seismic noise during training. A systematic experimental design with unified hyperparameters is employed to rigorously compare self-supervised and supervised strategies. The results demonstrate that synthetic Gaussian noise is ill-suited for real seismic denoising. In contrast, the proposed method effectively fine-tunes models without requiring clean labels, exhibiting high efficiency, strong generalization, and architecture independence across multiple network backbones, thereby significantly improving denoising performance.

Abstract

Self-supervised learning (SSL) has emerged as a promising approach to seismic data denoising as it does not require clean reference data. In this work, the deployment of the Noisy-as-Clean (NaC) method was evaluated for real seismic data denoising under controlled conditions. Two independent seismic acquisitions, each comprising noisy and filtered data, were organized into four real datasets. The NaC SSL method was adapted to add real noise to the noisy input, controlled by a parameter. An experimental protocol with ten experiments was designed to compare different strategies for deploying the NaC SSL method with the supervised learning baseline, using identical network topology and hyperparameters. The models were evaluated in terms of denoising performance, computational cost, and generalization capability. The results show that the synthetic additive white Gaussian noise (AWGN) is inadequate for the denoising of seismic data within the NaC method, and performance strongly depends on the compatibility between the injected and actual noise characteristics. Furthermore, both the characteristics of the seismic data and the noise level influence the performance of the model. Self-supervised fine-tuning on test data has improved SSL performance, whereas no such gain was observed for fine-tuning of supervised models. Finally, NaC has shown to be a simple, effective, and model-independent method that offers a feasible solution for the denoising of real seismic data.

Problem

• seismic data denoising
• self-supervised learning
• real noise
• Noisy-as-Clean
• data-driven denoising

Innovation

• Self-supervised learning
• Noisy-as-Clean
• Seismic denoising
• Real noise injection
• Model-independent denoising