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

Seismic data contains valuable information about the subsurface, but also contains noise that can be distracting. Denoising is an important step in seismic data processing. We used a U-Net-shaped encoder-decoder network with ResNeXt blocks and added self-attention to the deeper layers of the network. Additionally, we implemented attention gates as a second mechanism in order to attend to global features next to the local ones. We evaluated the impact of attention on machine-learning-based processing of seismic data by comparing results obtained without attention, with single attention, and dual attention. The results indicated that dual attention yielded improved denoising results compared to single attention, and single attention produced better results than without attention. Our proposed method resulted therefore in better preservation of the desired seismic signal.