Five-Dimensional Intelligent Denoising via Transformer

Deep learning for high-dimensional data has become a major concern in many fields as data scales continue to expand. With the rapid development of exploration seismology, the amount of seismic exploration data is growing rapidly. The 5-D seismic exploration data consist of two spatial coordinates for both shot and receiver points, along with one temporal coordinate. Moreover, with the complicacy of underground structural features and surface conditions, exploration data processing faces difficulties originating from low signal-to-noise ratios. Therefore, direct processing of 5-D data is necessary to fully leverage their structural features. However, both traditional methods and CNN approaches are not directly feasible for 5-D seismic data, making advancements in denoising technology a pressing issue. In this study, we introduce the Transformer model with 5-D positional embedding to fully exploit the structural information in higher dimensional space, achieving effective 5-D seismic denoising. With the help of transfer learning, the method can denoise the real data well. This framework can also be applied to high-dimensional seismic data interpolation and reconstruction through a pretraining-fine-tuning procedure.