No 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.

No abstract

During seismic data acquisition, receivers with finer trace intervals can be distributed to record seismic signals. The distance between neighbouring shots, on the other hand, can be significant, resulting in regularly missing shots in the common receiver gathers, which may cause significant spatial aliasing and reduce the precision of later seismic imaging. Traditional interpolation algorithms have several issues and limitations for seismic data with spatial aliasing due to prior assumptions and human–computer interactions. As a result, we present a new deep learning method that includes the generation of adaptive training data and the design of a consistent convolution kernel. Deep learning has the ability to characterize seismic data in a nonlinear manner that is ideal for accurate seismic interpolation. Because the common shot gathers and common receiver gathers have similar features due to the spatial reciprocity of Green's function of the wave equation if all shots have the same physical signatures, the common shot gathers with a refined trace interval are adaptively extracted as the training dataset to guarantee the interpolation performance in the common receiver gathers. With regularly subsampled data as input an improved U‐Net is created and trained to match the desired output, that is, the completed data. The trained network can be deployed to the test common receiver gathers to rebuild regularly missing shots. We develop a novel consistent convolution kernel to ensure high accuracy of missing shot reconstruction while accounting for differences between prestack unmigrated seismic data and images. Using numerically subsampled synthetic and field data, the effectiveness and validity of the developed consistent convolution kernel and the upgraded U‐Net with adaptive training data for missing shot interpolation are demonstrated.

No abstract