Seismic Data Reconstruction based on Attention U-net and Transfer Learning

Abstract

In field seismic exploration, missing seismic traces is inevitably encountered due to the constraints of the exploration environment and equipment. Thus, seismic data reconstruction is essential for seismic exploration analysis. In this paper, an attention-based U-net (AU-net) architecture is proposed by incorporating the attention mechanism into the U-net structure to address seismic data reconstruction challenges. The network incorporates both channel and spatial attention modules. Experiments comparing the reconstruction results of U-net, DnCNN, curvelet, and AU-net demonstrate the robustness of deep learning methods in handing increased missing trace percentages. The AU-net achieves superior reconstruction effect on contiguous missing regions, and exhibits better generalization through transfer learning experiments. To facilitate network training, this paper introduces a novel data-slicing technique to split and merge rectangular data for reconstructing shot records. This approach can be applied to any network and has high practical value.