Método de aprendizado profundo autossupervisionado para deblending de dados sísmicos marítimos

Abstract

Conventional seismic acquisition requires a time interval between shots long enough to record all events of interest and avoid interference between signals from successive shots. This causes a limitation both from the spatial data sampling point of view and the operational one. In order to overcome these limitations, the concept of blended seismic acquisition, also known as simultaneous sources acquisition, was introduced. In this technique, two or more shot points are fired with a shorter time between shots, allowing interference between data from different sources. However, the benefits brought by this technique are associated with seismic interference noise (blending noise) that must be treated in data processing. This seismic processing step is known as deblending. Recently, deep learning based techniques have been applied to several problems in geophysics. Since in a blended seismic acquisition there is no blending noise-free data available, the use of deblending methods involving supervised learning is limited, once this kind of method needs clean data to be used as targets for training. For this reason, this work proposes the use of the Trace2Trace self-supervised method for marine seismic data deblending. The method was implemented with different convolutional neural networks architectures such as U-NET, No-downscale CNN (NDCNN) and RES-NDCNN. The method was validated by comparing the results obtained in terms of the signal-to-noise ratio (SNR) and NRMSE metrics with the supervised deblending method as well as the sparse inversion deblending method, using the 2D windowed Fourier transform and the FISTA algorithm. In this stage, the effects of the architecture, the objective function and the post-processing procedure were evaluated in experiments performed on both synthetic and field datasets. The proposed method using the RES-NDCNN architecture showed better results than the other ones evaluated. After choosing the architecture, the effect of the random time shifts added to the time intervals between shots (dither times) on the results was analyzed, as well as the network’s ability to generalize when applied to a database other than the one used for training. The results obtained by the proposed method are comparable to those reported in literature under similar conditions and outperform the sparse inversion deblending method for dither times in the [-0.5, 0.5] s range.