Deep learning-based artificial bandwidth extension: Training on ultrasparse OBN to enhance towed-streamer FWI

Abstract

The lack of seismic low frequencies in towed-streamer data is known to have an outsized detrimental effect on advanced velocity model building techniques such as full-waveform inversion (FWI). Since seabed acquisition records ultralow frequencies (1–4 Hz) with high signal-to-noise ratio, this presents an opportunity to learn, in a supervised machine learning fashion, a bandwidth extension function to enrich towed-streamer data with low frequencies. We use recent advances in training deep neural networks to develop a novel method for learning low-frequency reconstruction from an ultrasparse set of ocean-bottom nodes (OBNs). This bandwidth extension is tested on two large field data sets (from an OBN survey and a wide-azimuth towed-steamer survey) acquired over a complex-shaped salt region in the Gulf of Mexico. The reconstructed low frequencies, although not perfect, enable FWI to more effectively correct the shape of salt bodies and result in improved subsalt imaging. Well-tie analysis shows an improvement in phase stability around the wellbore and a fit to within half a cycle at reservoir level. This work links together towed-streamer and seabed acquisitions, providing a cost-effective solution to help offshore seismic exploration with higher-quality low frequencies.