STUGAN: An Integrated Swin Transformer-Based Generative Adversarial Networks for Seismic Data Reconstruction and Denoising

The collection process of seismic data is often affected by terrain conditions and human factors, resulting in spatial gaps, undersampling, and random noise in the collected seismic trace set, which hampers the subsequent seismic data processing and interpretation. Existing seismic data reconstruction and denoising algorithms based on deep learning mostly use local similarity learning to obtain the distribution of missing traces or noisy data and lack nonlocal similarity feature mining, which cannot effectively extract effective features of seismic data under the mixed interference of several missing traces and strong noise. This poses challenges to simultaneous reconstruction and denoising tasks. This study introduces swin transformer (ST) into seismic data processing. Combined with the generative adversarial network (GAN) structure, a new simultaneous reconstruction and denoising model, ST-united GAN (STUGAN), is proposed. First, an ST module is used to replace the traditional convolution module in GAN, and its self-attention mechanism captures the correlation between local and nonlocal seismic data features to improve the model feature extraction ability. Second, the generator is constructed based on U-Net, and the event features and texture information features of seismic data are extracted. Conditional constraints are incorporated into the discriminator to guide the gradient optimization direction of the generator. Finally, through synthetic and real data experiments, we demonstrate that STUGAN has a good recovery effect and is robust to missing seismic data in large gaps and strong noise interference.