Seismic random noise attenuation in the Laplace domain using deep learning

We attenuated random noise in Laplace-domain seismic wavefields using a modified U-net. Laplace-domain wavefields can be obtained by Laplace-transforming time-domain wavefields. Due to the damping in the Laplace transform, small-amplitude noises near the first arrival signal can severely contaminate Laplace-domain wavefields. Therefore, time-domain denoising is not sufficient for seismic data processing in the Laplace domain. We trained a modified U-net in a supervised manner to generate clean wavefields from noisy wavefields. Since Laplace-domain wavefields show exponential decay with respect to offset, we used the logarithmic representation of the wavefields to train the network. Numerical examples show that the deep-learning approach can attenuate random noise better than denoising using singular value decomposition.