Robust low frequency seismic bandwidth extension with a U-net and synthetic training data

Abstract

This work focuses on enhancing low frequency seismic data using a convolutional neural network trained on synthetic data. Traditional seismic data often lack both high and low frequencies, which are essential for detailed geological interpretation and various geophysical applications. Low frequency data is particularly valuable for reducing wavelet sidelobes and improving full waveform inversion (FWI). Conventional methods for bandwidth extension include seismic deconvolution and sparse inversion, which have limitations in recovering low frequencies. The study explores the potential of the U-net, which has been successful in other geophysical applications such as noise attenuation and seismic resolution enhancement. The novelty in our approach is that we do not rely on computationally expensive finite difference modelling to create training data. Instead, our synthetic training data is created from individual randomly perturbed events with variations in bandwidth, making it more adaptable to different data sets compared to previous deep learning methods. The method was tested on both synthetic and real seismic data, demonstrating effective low frequency reconstruction and sidelobe reduction. With a synthetic full waveform inversion to recover a velocity model and a seismic amplitude inversion to estimate acoustic impedance we demonstrate the validity and benefit of the proposed method. Overall, the study presents a robust approach to seismic bandwidth extension using deep learning, emphasizing the importance of diverse and well-designed but computationally inexpensive synthetic training data.