MSAACNN for intense noise suppression in DAS-VSP records.

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Distributed optical fiber sensing (DAS) is increasingly used in seismic exploration due to its acquisition and deployment advantages.
• DAS records often suffer from low signal-to-noise ratio (SNR) caused by intense background noise.
• Effective noise suppression in DAS data is crucial for seismic data processing.

Purpose of the Study:

• To address the challenge of intense background noise suppression in DAS records.
• To propose a novel deep learning model for enhancing DAS data quality.
• To improve the signal-to-noise ratio (SNR) of DAS seismic data.

Main Methods:

• A multiscale sparse asymmetric attention convolutional neural network (MSAACNN) was developed.
• The network utilizes dilated convolutions for expanded receptive fields and asymmetric convolutions for enhanced feature extraction.
• A pyramid attention module was incorporated to refine features and boost denoising performance.

Main Results:

• The MSAACNN effectively suppressed complex background noise in DAS records.
• Compared to traditional methods and standard CNNs, MSAACNN demonstrated superior denoising capabilities.
• Recovered signal components were clear and complete, with a significant improvement in SNR.

Conclusions:

• The proposed MSAACNN is a powerful tool for DAS background noise suppression.
• This deep learning approach significantly enhances the quality of seismic data acquired via DAS.
• The method offers a promising solution for improving SNR in seismic exploration using DAS technology.