Iterative deblending using MultiResUNet with multilevel blending noise for training and transfer learning

Abstract

Blended seismic acquisition has improved the efficiency of land and marine data acquisition significantly. Nevertheless, the consequent blending noise poses challenges for subsequent seismic imaging and inversion. So, deblending algorithms are being widely investigated. To improve the deblending performance and efficiency of traditional deblending algorithms, we propose a method that we call Multiresolution ResUNet (MultiResUnet) trained on datasets with multi-level blending noise. The trained MultiResUnet is optimal for iterative deblending. MultiResUNet combines the advantages of a residual learning network (ResNet) and U-net. We apply a scalar factor to the synthetic blending noise to construct a training dataset with multi-level blending noise contamination, which is designed to train and fine-tune MultiResUNet to detect weak blending noise and effectively retrieve signals in an iterative manner. The proposed method attenuates the blending noise iteratively, leading to an improved performance compared to the conventional curvelet transform-based thresholding algorithm or MultiResUNet trained on datasets with single-level blending noise. Since the training and fine-tuning of MultiResUNet happens only once up front, the application of the trained MultiResUNet is efficient. To demonstrate the performance of the proposed method, the improved deblending accuracy is verified through comparison on numerical examples. For field data applications, a transfer learning approach is adopted to generalize the MultiResUNet trained on synthetic blended data for accurate deblending.

Focal deblending: Marine data processing experiences

“…B. Wang et al. (2022) use iterative inversion and a multi‐resolution U‐Net to take advantage of the multiscale nature of seismic data. Deep neural network–based approaches can also be used for gradient denoising in iterative schemes, both in a supervised (K. Wang, Mao, et al., 2022; K. Wang & Hu, 2022) and in an unsupervised (K. Wang, Hu, et al., 2022) fashion.…”