Simultaneous‐source deblending using adaptive coherence‐constrained dictionary learning and sparse approximation

Abstract

The dictionary learning and sparse approximation method using the K‐singular value decomposition algorithm rely on the knowledge of the sparsity or noise variance as a constraint when it is used for data denoising. However, the determination of the sparsity or noise variance of seismic data can be tricky and sometimes unknown, especially in seismic field data. Thus, where the cardinality or the noise variance is not known, the intrinsic character of the relative coherence between the removed noise from noisy data and its learned dictionary is instead used as a constraint for the sparse approximation of simultaneous‐source seismic data. The dictionary learning is obtained using a modified orthogonal matching pursuit algorithm which uses coherence as a constraint and is referred to as coherence dictionary learning. The coherence dictionary learning is then adapted to handle the simultaneous‐source seismic data deblending. A blending structure with random time dithering of sequential source shooting is used to guarantee adequate randomness of the noise. Two‐dimensional overlap patches of the noisy data were extracted from the common receiver gather domain to train the dictionary and to determine the sparse representation of the signal. The method is tested on both synthetic and field data, and it shows adequate data recovery. Comparing the result of this method to the matching pursuit algorithm constrained by the signal sparsity and the noise variance reveals that our approach performs better at noise attenuation and yields a reasonable data recovery especially for strong seismic signal.

A multi‐data training method for a deep neural network to improve the separation effect of simultaneous‐source data

“…The Marmousi model data contain a large number of normal faults and steep dip angle structures, and the lateral velocity changes drastically, which is consistent with the actual complex underground medium conditions (Evinemi & Mao, 2021). It is a complex geological model used by the petroleum industry for method testing and the finite‐difference algorithm is carried out to obtain acoustic data.…”

“…We test the performance of our proposed method using field marine data acquired with conventional shooting and published by the PGS company (Evinemi & Mao, 2021). The numbers of shots and receivers of the field data are both 256, and the temporal sampling interval is 4 ms. We select 50 records to make the training data sets for the DNNGT with parameters $${{\bm{\theta }}}^D$$, and for the DNNMDT with parameters $${{\bm{\theta }}}^E$$.…”

“…We test the performance of our proposed method using field marine data acquired with conventional shooting and published by the PGS company (Evinemi & Mao, 2021). The The DNNGT method with parameter 𝜽 𝐷 and the DNN-MDT method with parameter 𝜽 𝐸 are, respectively, used to deblend the blended data.…”

“…(2019) proposed transforming seismic data to the frequency–wavenumber–wavenumber (FKK) domain as an effective method of simultaneous‐source data separation. Using shaping the regularization framework, some scholars have also considered using coherence dictionary learning (Evinemi & Mao, 2021) and compressive sensing (Kumar et al., 2015, 2018; Mansour et al., 2012; Wang & Geng, 2020) methods to complete the simultaneous‐source data separation. While achieving a certain separation effect, the dictionary learning method is computationally expensive.…”

An improved algorithm of segmented orthogonal matching pursuit based on wireless sensor networks

“…The StOMP algorithm determines the atoms through a threshold, and the input parameters do not include the signal sparsity K. Compared with the ROMP algorithm, the StOMP algorithm has higher computational efficiency and reconstructed signal accuracy. 21 However, the StOMP algorithm still has a problem, which is the setting of the threshold. The threshold cannot be too large or too small.…”