Ground-roll attenuation in high-dimensional domain using multi-scale decomposition and attention mechanism network

Abstract

Ground-roll attenuation is a critical step for enhancing the signal-to-noise ratio of seismic data, serving as an essential prerequisite for achieving precise subsurface imaging. Due to the overlap of ground-roll and useful signals in both the t-x and frequency domains, coupled with the high amplitude of ground-roll, existing methods often struggle to remove strong ground-roll without damaging useful signals. To address this issue, we propose a ground-roll attenuation method that integrates multi-scale decomposition and attention mechanism network, reframing the ground-roll attenuation problem in the t-x domain as the prediction of ground-roll coefficients in a high-dimensional transform domain. Given the differences in frequency and propagation direction between ground-roll and useful signals, we employ the Non-Subsampled Shearlet Transform, which provides multi-scale and multi-directional decomposition capabilities, to decompose seismic data containing ground-roll into multiple subbands, thereby reducing the overlap between ground-roll and useful signals. This enables the network to more easily and accurately extract ground-roll while retaining useful signals. Furthermore, to effectively extract ground-roll coefficients from subbands at different scales, we propose a Multi-Scale Attention Network. The network features a multi-branch structure with convolutional kernels of various sizes, enabling the capture of waveform features across different scales. An attention mechanism is then used to select and fuse feature maps from different branches, further enhancing the network's ability to capture information across scales. Experimental results both on synthetic and field data demonstrate that, the proposed method outperforms a conventional method and two advanced deep learning methods, achieving superior ground-roll attenuation while better preserving useful signals.