Ground-roll attenuation using dual-model self-supervised selective learning with blind horizontal convolutional neural networks

Ground-roll is a coherent noise that we inevitably encounter during seismic data acquisition on land. It broadly conceals the reflected wave signals, reducing the signal-to-noise ratio (SNR) of data. To attenuate ground-roll, various machine learning techniques have been studied. Recently, the label-free self-supervised learning-based techniques have become actively studied, and the ground-roll attenuation method using the two U-Nets and the loss function combining the Fourier and misfit losses has shown high accuracy. However, this method suffers from incomplete separation of ground-roll from desired signals, which is caused by the identity mapping problem of U-Net, and has instability due to the loss function. To mitigate this problem, we propose using the blind horizontal network (BHN) and dual-model self-supervised selective learning (dSSSL). BHN is designed by removing horizontal pixels in the vertically aligned receptive fields to prevent the identity mapping and effectively separate ground-roll from seismic data. For dSSSL, we use the output image from the first network and its residuals with respect to the input to redistribute ground-roll and desired signals. The synthetic data experiment shows that the proposed ground-roll attenuation method improves the accuracy and convergence stability. For the synthetic data, we also investigate the effects of user-defined parameters such as weighting factors and frequency constraints. Furthermore, we compare our method with the widely used f-k filter for the field data acquired in Pohang, South Korea, which indicates that our approach has a performance close to f-k filtering.