Efficient seismic data denoising via multi-scale attention network with depthwise separable and residual dilated...

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• High signal-to-noise ratio (SNR) is crucial for accurate seismic data processing.
• Complex noise interference poses a significant challenge in seismic signal denoising.

Purpose of the Study:

• To introduce MSARDNet, a novel network for seismic signal denoising.
• To address the challenge of complex noise interference in seismic data.

Main Methods:

• MSARDNet employs a multi-level feature processing architecture with encoder, ResDC, and decoder modules.
• Depthwise separable convolution and enhanced channel attention are used for efficient feature extraction and noise suppression.
• Dilated convolutions and residual connections in the ResDC module facilitate multi-scale feature extraction and preserve signal integrity.

Main Results:

• MSARDNet achieved SNR improvements of 15.94 dB, 12.03 dB, and 4.83 dB over U-Net, DnCNN, and DeepDenoiser, respectively.
• The network effectively filters complex background noise while preserving useful seismic signals.
• Comparative experiments on synthetic and field data validated the superior performance of MSARDNet.

Conclusions:

• MSARDNet demonstrates superior performance in eliminating complex noise compared to existing methods.
• The proposed network holds significant potential for broad application in seismic data processing and analysis.