Deep learning for near-surface waves separation in the frequency–slowness domain

Ivan Sanchez, Daniel O. Trad

Near-surface waves often dominate land seismic records and obscure body-wave information. To address this problem, we propose a deep learning workflow based on a U-Net architecture trained in the frequency–slowness domain. Training datasets are generated with the Partitioned Domain Method (PDM), which produces total and near-surface wavefields from finite-difference elastic simulations of models with complex topography and heterogeneous near-surface conditions. Polarization analysis is used to compute the semi-major and semi-minor axes of the spectral ellipses, and two U-Net autoencoders are trained to predict filter masks that highlight near-surface wave patterns in the (f,p) domain. The predicted masks enable reconstruction and subtraction of near-surface energy from multicomponent seismic data, improving the visibility of deeper reflections. Results using synthetic data from the SEAM Foothills Phase II model show that the proposed workflow effectively attenuates near-surface waves in 3C seismic data.