物探化探计算技术2026，Vol.48Issue(4)：499-513,15.DOI:10.12474/wthtjs.20260126-0002
融合Transformer与多尺度注意力机制的地震数据重建方法
A seismic data reconstruction method integrating Transformers and multi-scale attention mechanisms
摘要
Abstract
During seismic data acquisition in mountainous regions,constraints such as surface conditions and economic costs often result in spatial discontinuities caused by missing traces in the collected seismic data,which severely compromise the accuracy of subsequent seismic reflection imaging and interpretation.To address the limitations of traditional interpolation methods—which suffer from low reconstruction accuracy in complex structural areas—and the difficulty of conventional Convolutional Neural Networks(CNNs)in capturing the global long-range dependencies of seismic wavefields,this paper proposes an improved U-Net3+seismic data interpolation method that integrates the Transformer architecture with a multi-scale attention mechanism(Convolutional Block Attention Module,CBAM).This method uses U-Net3+as its base architecture and employs full-scale skip connections to achieve deep integration of multi-scale features.A Transformer module is introduced at the end of the encoder to effectively model global contextual information in seismic signals via self-attention mechanisms,compensating for the limited receptive field of convolutional operations.Additionally,CBAM attention modules are embedded in the skip connections to enhance effective wavefield features while adaptively suppressing background noise.Experimental results on two-dimensional synthetic seismic records and real seismic data demonstrate that this method can still reconstruct seismic records with high quality even under a 50%random missing rate.Compared to traditional U-Net-based methods,the method proposed in this paper demonstrates significant advantages in quantitative metrics:the Peak Signal-to-Noise Ratio(PSNR),used to evaluate signal reconstruction error,improves by up to 31.08%(5.89 dB),and the Structural Similarity Index Measure(SSIM),used to assess the similarity of spatial features such as wavefield in-phase axes,improves by up to 30.20%.This method significantly enhances the wavefield fidelity of seismic data and can more clearly reconstruct complex geological features such as faults and unconformities,demonstrating strong robustness and practical value.关键词
地震数据插值/深度学习/U-Net3+/波场重建Key words
seismic data interpolation/deep learning/U-Net3+/wavefield reconstruction分类
天文与地球科学引用本文复制引用
刘芳,李怀良,杨李欣..融合Transformer与多尺度注意力机制的地震数据重建方法[J].物探化探计算技术,2026,48(4):499-513,15.基金项目
四川省青年科学基金1001-1749B1001-1749类(No.2026NSFSC1134) （No.2026NSFSC1134）
中国国家自然科学基金面上项目(No.42474199) （No.42474199）
四川省青年科技创新团队项目(No.2026NSFSCZY0058) （No.2026NSFSCZY0058）
