Extrapolating low-frequency prestack land data with deep learning

Missing low-frequency content in seismic data is a common challenge for seismic inversion. Long wavelengths are necessary to reveal large structures in the subsurface and to build an acceptable starting point for later iterations of full-waveform inversion (FWI). High-frequency land seismic data are particularly challenging due to the elastic nature of the Earth contrasting with acoustic air at the typically rugged free surface, which makes the use of low frequencies even more vital to the inversion. We propose a supervised deep learning framework for bandwidth extrapolation of prestack elastic data in the time domain. We utilize a Convolutional Neural Network (CNN) with a UNet-inspired architecture to convert portions of band-limited shot gathers from 5-15 Hz to 0-5 Hz band. In the synthetic experiment, we train the network on 192x192 patches of wavefields simulated for different cross-sections of the elastic SEAM Arid model with free-surface. Then, we test the network on unseen shot gathers from the same model to demonstrate the viability of the approach. The results show promise for future field data applications.