# Multi-scale dual-path attention network for seismic background noise attenuation

## TL;DR

This paper introduces a new deep learning network for reducing background noise in seismic data, improving the accuracy of geological analysis.

## Contribution

The novel MSDPA-Net uses multi-scale features and dual-path attention to enhance seismic noise suppression.

## Key findings

- MSDPA-Net outperforms traditional algorithms in suppressing complex seismic noise.
- The network effectively handles non-Gaussian and nonlinear noise characteristics.
- Experiments on simulated and field data show improved denoising performance.

## Abstract

The background noise in seismic records severely interferes with the extraction of effective reflection events, particularly in complex exploration environments such as deserts. The non-Gaussian and nonlinear characteristics of background noise further exacerbate the difficulty of noise reduction, impacting the accuracy of subsequent processing such as inversion and migration. In recent years, deep learning has demonstrated excellent performance in the suppression of complex seismic noise, exhibiting notable advantages over traditional denoising algorithms. However, traditional deep learning networks often focus solely on feature extraction at a single scale, which proves inadequate when handling complex and variable seismic data. To address the aforementioned problem, we propose a Multi-scale Dual-path Attention Network (MSDPA-Net) aimed at enhancing denoising effectiveness by fully leveraging the multi-scale features of seismic data. MSDPA-Net employs a multi-scale strategy for preliminary feature extraction, followed by a dual-path attention module to discriminate between signal and noise. Subsequently, it utilizes a feature interaction structure for reinforcement learning, culminating in effective information fusion and reconstruction through a reconstruction module. Experiments on simulated and field seismic data demonstrate that MSDPA-Net exhibits remarkable performance in suppressing complex seismic noise compared to traditional denoising algorithms and typical deep learning networks.

The online version contains supplementary material available at 10.1038/s41598-025-25446-x.