WAHNet: A Wave-Attention Hybrid Network for Surface Wave Suppression in Seismic Data

Abstract

Noise in seismic data includes random noise and coherent noise. Surface waves, as typical coherent noise, are characterized by low-frequency, low apparent velocity and high amplitude. Its physical propagation characteristics are similar to some effective signals, which leads to the difficulty of distinguishing them in the time-frequency domain. Convolutional neural networks have been increasingly used for surface wave suppression due to their adaptive feature extraction. However, they struggle in separating surface waves from effective signals with their overlapping frequency range. In this paper, we propose a wave-attention hybrid network (WAHNet) to handle this issue. WAHNet has two types of new modules: (1) wavelet-gated enhancement module extracts and enhances the time-frequency domain features of seismic signals by combining wavelet convolution with a gating mechanism and (2) adaptive wavelet attention module introduces a joint attention mechanism in the time-frequency domain to alleviate the time-frequency aliasing problem of surface waves and effective signals. The experimental results show that, compared with the five popular methods, WAHNet can suppress surface waves more effectively while retaining richer effective signals.