.png){kind=icon}
.png){kind=icon}
Return
Low-Frequency Extrapolating Transformer
.png){kind=icon}
PRE
AI
DOI：10.1111/1365-2478.70194.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The low-frequency absence in seismic data is likely to influence the vertical resolution, the imaging for complex structures and the convergence of full-waveform inversion (FWI). The extrapolation task of absent low-frequency components is a strong non-linear problem, which can be solved by deep learning (DL) owing to its strong capability of non-linear mapping. Meanwhile, low-frequency extrapolation can be regarded as a global waveform transformation, in which global features are beneficial for improving the accuracy and generalization of DL-based low-frequency extrapolation methods. Therefore, we propose a Transformer-based network oriented for capturing more global features, called low-frequency extrapolation Transformer (LFET), to extrapolate the absent low-frequency components of seismic data. To our best knowledge, this is the first attempt to apply the Transformer model to the task of low-frequency extrapolation. This proposed LFET is composed of eight successively connected Transformer blocks (TBs) which can extract contextual and global information by using the core operation: multi-head self-attention (MSA). Moreover, we introduce the shifted window partitioning strategy into LFET, thus alleviating the computational burden almost without influencing the capture of global features. In the experimental part, we use several synthetic and field records to test the effectiveness of LFET. Our method exhibits better extrapolation performance than three convolution-based competitive methods as well as lower training and testing time-cost. Furthermore, in the discussion part, a benchmark model is used to preliminarily demonstrate the potential application of LFET in improving the convergence of FWI.
Keywords:
deep learning
low-frequency extrapolation
signal processing
Transformer
Journal
G
IF:
1.8
Papers:
110
Citations:
6.0K