Suppression of swell noise via dip-aligned self-supervised learning

Ji Li, Daniel O. Trad

Recently, self-supervised deep learning methods have gained popularity in seismic data denoising due to their ability to train without clean labels, which are often unavailable in real seismic applications. However, most existing approaches are limited to suppressing random noise or noise coherent only along the time axis. StructBS, a recently proposed self-supervised denoising network, has shown effectiveness in handling temporally coherent noise. In this work, we improve StructBS by integrating a self-attention module to capture long-range spatial dependencies, enabling the network to better suppress noise that is coherent across multiple traces. In addition, we introduce a dip-aligned gradient constraint that penalizes variation along the estimated dip direction, effectively guiding the denoising process to preserve geologically meaningful structures. This combined framework targets partially coherent noise, noise that is neither purely random nor fully coherent, commonly observed in field data. We evaluate the proposed method on the case of swell noise in marine seismic data. Quantitative and visual comparisons show that our method outperforms existing self-supervised networks and traditional denoising techniques in suppressing strong, partially coherent noise while preserving useful seismic signals and structural continuity.