A Self-Attention Enhanced Encoder-Decoder Network for Seismic Data Denoising

While seismic data contains an abundance of useful information about the subsurface, it is also contaminated with disturbing recorded energy, coherent and random noise. A crucial step in seismic data processing, therefore, is denoising. In this work, we use a U-Net-based encoder-decoder network that uses ResNeXt blocks rather than traditional convolutions as a denoising tool. Furthermore, we add self-attention to the ResNext blocks in the deeper part of the neural network architecture. We compare the results with those obtained with the same network, but without attention, in order to investigate, whether self-attention improves machine-learning-based processing of seismic data. The supervised training with the implemented attention mechanism leads to improved denoising results than without attention. The deleted noise contains less primary energy and thus leads to a better conservation of the desired seismic signal.