Application of Deep Learning for Low-Frequency Extrapolation to Marine Seismic Data in the Sadewa Field, Kutei Basin, Indonesia

Abstract

The success of full-waveform inversion (FWI) modeling depends on many factors, i.e., initial model, source wavelet, and low-content. Low-frequency seismic data is extremely important for FWI as it allows for accurate sub-surface imaging and increasing the resolution of sub-surface features, while also providing information on overall sub-surface geological features. However, the lack of low-frequency data can lead to cycle-skipping which could negatively affect the results of the inversion process. To resolve this issue, low-frequency data can be reconstructed using artificial intelligence or deep learning algorithms. In this study, we utilized a convolutional neural network (CNN) algorithm to automatically extrapolate the low-frequency data from band-limited Common Shot Gather (CSG) seismic data in time domain without any preliminary processing steps. The CNN model was tested and validated with synthetic seismic data to select the best model. The best model was then applied to marine seismic data in the Kutei Basin, East Kalimantan, Indonesia. The results demonstrate that deep learning has been accurately applied to extrapolate low frequencies in marine seismic data with an RMS error below 1% and that it produces results that are in line with the broadly established in conventional seismic processing methods (i.e., deghosting). Therefore, our study can be a powerful and effective alternative method to estimate the low-frequency content in the seismic reflection data of other areas.