基于改进U-Net模型的地震数据去噪方法研究

Research on Seismic Data Denoising Method Based on Improved U-Net Model

摘要: 鉴于传统U-Net模型在地震数据去噪中存在特征提取不足的问题，本文引入U-Net++模型，通过嵌套式分层密集连接和优化跳跃连接，增强局部细节与高层语义特征的融合，改善复杂噪声的抑制效果，并融合CBAM注意力机制构建CBAM-U-Net++模型，利用通道注意力(CAM)与空间注意力(SAM)建立语义依赖，提高图像特征提取的精准性与鲁棒性。实验结果表明，CBAM-U-Net++模型在合成数据和实际地震记录上均表现出优异的噪声去除和特征保留能力，在复杂噪声环境下优势明显。与传统U-Net及U-Net++模型相比，CBAM-U-Net++模型的PSNR分别提高了6.3 dB和4.08 dB，SSIM提升了0.12和0.09。

Abstract: In view of the limitations of the conventional U-Net model in feature extraction for seismic denoising, this study introduces the U-Net++ model. U-Net++ enhances the fusion of local details and high-level semantic features through nested, hierarchical dense connections and optimized skip connections, thereby improving the suppression of complex noise, this study integrates Convolutional Block Attention Mechanism to construct the CBAM-U-Net++ model. By employing channel attention (CAM) and spatial attention (SAM) to establish semantic dependencies, the model achieves enhanced extraction precision and robustness. Experimental results on both synthetic and real seismic datasets demonstrate that the CBAM-U-Net++ model exhibits superior denoising and feature preservation capabilities, particularly in complex noise environments. Compared with the conventional U-Net and U-Net++ models, the CBAM-U-Net++ model achieves improvements of 6.3 dB and 4.08 dB in PSNR, and enhancements of 0.12 and 0.09 in SSIM, respectively.