A Hybrid CNN-Transformer Network for Linear Noise Suppression in 3D Seismic Data

Abstract 3D seismic data acquisition, processing, and analysis have become increasingly prevalent in modern geophysical exploration. However, various types of coherent noise, especially linear noise, can cause serious interference and effective noise reduction techniques are critical. Most deep learning-based denoising methods rely on converting 3D seismic volumes into 2D slices and then reassembling them, which disrupts the spatial continuity and structural integrity of the data. To address this issue, we propose a novel linear noise suppression network capable of directly processing 3D seismic volumes, effectively preserving spatial features and continuity. Our proposed approach, the 3D Deformable Convolution-Transformer U-Net (3D-DCTU) network, integrates 3D deformable convolutional layers with a Transformer architecture. The deformable convolutional layers adaptively adjust sampling locations through learned offsets to capture spatial variations in seismic data, while the Transformer's self-attention mechanism captures long-range dependencies. This combination enables both the extraction of local features and the modeling of global information. To enhance perceptual quality of the denoised results, the network incorporates the gradient difference loss and mean squared error as optimization targets. We evaluated the proposed network using both synthetic data and field data from a region in western China. The results suggest promising denoising performance, particularly for linear noise attenuation, indicating a viable technical pathway for 3D seismic data denoising.