Answers are grounded in this paper and cite their sources.
Data-driven methods provide an end-to-end framework for seismic noise attenuation, but the limited receptive field of conventional convolutional neural networks (CNNs) restricts their ability to model long-range dependencies in seismic data. Their generalization is further limited by the distribution mismatch between synthetic training data and field records. To address these challenges, we propose a local-orthogonality-constrained Swin Transformer convolutional network (LOSTCG). The core STCG module combines convolution-based local feature extraction with attention-based adaptive feature weighting, enabling effective multiscale representation of seismic events. In addition, a local orthogonality term is incorporated into the loss function to reduce the leakage of coherent signals into the estimated noise and improve robustness to complex field noise. A diverse synthetic training dataset is also constructed and augmented with field noise to narrow the distribution gap between synthetic and field data. Experiments on synthetic data and low-signal-to-noise-ratio field records demonstrate that LOSTCG achieves better overall performance in noise attenuation and signal preservation than the f–k method, singular value decomposition (SVD), ResUNet, and SCRN. These results indicate that the proposed method generalizes well to field data and has considerable potential for practical seismic data processing.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: