Research On Blended Seismic Data Processing Method Based On Sparse Representation

Seismic data acquisition is the first step in seismic exploration,but traditional data acquisition takes a long time and the huge cost.Seismic acquisition based on（nearly）simultaneous source makes up for the shortcomings of traditional seismic data acquisition.It allows the simultaneous shooting of seismic sources or with smaller temporal,and reduces acquisition cost^[1],but also produces blending noise.During the data collection process,other factors may be affected,resulting in the traces missing,such as the surrounding environment and poor contact of the detector.Since the data are blended,in order to avoid the introduction of additional noise during seismic data reconstruction,the blending noise is removed first,and then the seismic data reconstruction is performed,and we obtain the clean single shot recorded by the traditional.The sparse transformation of signals has been in the ascendant in the past 20 years,and its great advantage is that it can better represent the most signal information with the least amount of data.After the signal is sparsely transformed,the sparse coefficients are denoised and reconstructed in the sparse domain,the effective signal is separated,and then converted back to the original domain,and the processing of the data is completed.Based on sparse transformation,this paper studies the reconstruction method of seismic data and the separation of blended data.In the acquisition principle part,this paper introduces the basic composition concept,the representation of matrix and image of multi-source acquisition data.In the acquisition records of different domains,the common receiver domain is applied to the subsequent processing;in the sparse representation theory,the basic fixed base is introduced,including the formula representation and examples of each sparse theory,these fixed-base sparse transforms will be applied to the subsequent reconstruction and separation comparison;the Radon transform method of this paper is introduced in detail.It contains three main Radon transforms,which illustrate the scope of use for each method and their respective advantages.The hyperbolic Radon transform is applied in the subsequent processing;the dictionary learning theory is introduced in detail,including the K-SVD dictionary and the Alternating Direction Method of Multipliers（ADMM）.The ADMM algorithm was used in the subsequent Seismic data reconstruction,and the K-SVD dictionary was used in the seismic blended data separation.Aiming at the problems of dictionary learning and ADMM（alternating direction method of multipliers）based methods,we propose a compressed sensing based seismic data restruction method with dictionary learning and ADMM.In this case,the new method can be separated into three steps.Firstly,we use a learned dictionary to sparsely represent the seismic data.And then,a reasonable measurement matrix can be designed according to the missing extent of the seismic data traces.Finally,the L1norm constrained model is solved by ADMM to obtain the restructed seismic data.In addition,we establish a compressed sensing based technical implementation flow of seismic data interpolation with dictionary learning and ADMM.Compared with the compressed sensing based restruction methods of fixed basis,the learning dictionary can provide a better sparse representation for the seismic data adaptively.And compared with the conventional restruction algorithms,,such as curvelet,ADMM can recover seismic data more accurately.The synthetic data and real data restruction experiments indicate that the seismic data restructed by dictionary learning and ADMM based on compressed sensing has higher signal-to-noise ratio（SNR）than that restructed by Fixed base and OMP.The authors propose an approach which can separate blended seismic acquisition data and this method is based on the seismic sparse inversion.Meanwhile,the proposed approach updates the dictionary atoms via K-time iterative singular value decomposition（K-SVD）in the Radon domain:the separation of blended seismic data can be regard as the sparse inversion in the case of simultaneous sources.So in the common detector data,the sparse representation of the events are convergent when they transform into Radon domain.After that,the filters and wise-block dictionary learning are applied in the blend data to represent the seismic data.Finally,we fix the updated dictionary and calculate the sparse coefficients to achieve the separation.Through the simulation and field data experiments,we find that separation method of the proposed approach is more accurate than median filter and wavelets transform.So that the proposed approach will be applied in the practice.