PSD-Net: A Physics and Structure Guided Dual-Stream Transformer for 3-D Seismic Denoising
Abstract
3-D seismic data denoising is a fundamental basis for geophysical imaging and interpretation. Although deep learning methods have made significant progress in random noise attenuation in recent years, existing methods are still prone to smoothing and blurring key geological structures such as faults and stratigraphic boundaries, making it difficult to balance noise suppression and structural fidelity. To address this, this letter proposes a dual-stream 3-D Transformer denoising network jointly guided by physics- and structure-tensor dual-guided dual-stream Transformer denoising network (PSD-Net). The network adopts a decoupled dual-stream architecture consisting of a noise separation stream (NSS) and a feature restoration stream (FRS): the NSS branch utilizes a 3-D structure tensor to construct a structure-guidance map, which comprehensively characterizes local directional anisotropy and structural energy, and subsequently learns a noise-aware spatial mask to achieve explicit modeling of noise regions; meanwhile, the mainstream FRS integrates a physics prior constraint module that applies spatial smoothness constraints to features via a 3-D Laplacian operator to maintain spatial continuity consistent with valid signals, while combining spatially adaptive amplitude modulation to enhance local feature representation. The features from both streams adaptively interact and fuse at each encoding scale through dual-stream fusion blocks. Experiments on both synthetic and real 3-D seismic datasets demonstrate that PSD-Net outperforms current 3-D denoising methods in terms of random noise suppression and complex geological structure preservation.