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

Because a high signal-to-noise ratio (SNR) is critical in enhancing the accuracy of subsequent processing, noise reduction remains a pivotal challenge in seismic signal processing, especially for complex noise interference scenarios. To address this, we propose a novel network termed MSARDNet for seismic signal denoising, with its novelty in a multi-level feature processing architecture integrating innovative modules. Specifically, the architecture comprises an encoder module, a ResDC module, and a decoder module. The encoder integrates depthwise separable convolution to efficiently extract local features with reduced parameters and improved computational efficiency. Prior to pooling, an enhanced channel attention mechanism is incorporated to adaptively emphasize critical features and suppress noise based on their importance, further boosting denoising efficacy. The ResDC module combines dilated convolutions for multi-scale receptive field expansion and residual connections for gradient optimization, ensuring feature integrity and facilitating joint extraction of multi-scale features and global context. This design effectively mitigates information loss, preserves useful signals, and enhances SNR by filtering complex noise. Through comparative experiments on synthetic and field data using denoising methods such as U-Net, DnCNN, and DeepDenoiser, MSARDNet achieved signal-to-noise ratio improvements of + 15.94 dB, + 12.03 dB, and + 4.83 dB over these methods, respectively. The results demonstrate that MSARDNet outperforms existing methods in eliminating complex background noise and holds broad application prospects in seismic data processing.