Research On Denoising Method Of Multi-dimensional Seismic Data Based On Sparse Representation

Due to the influence of acquisition mode and acquisition environment in the process of seismic data acquisition,problems such as noise pollution and poor data quality will occur in seismic data acquisition,which will affect the processing results of subsequent seismic data.therefore,it is of great significance to remove and reconstruct the original seismic data collected in the field,improve the signal-to-noise ratio and enhance the resolution.Sparse representation is a popular data representation method in recent years,which can well represent the data information of seismic signals.this method has strong data processing ability and feature extraction ability.This paper evaluates the denoising and reconstruction effects of the sparse representation denoising method based on fixed basis function feature extraction and the traditional dictionary learning feature extraction sparse representation denoising method based on data-driven adaptability.in view of the defects in the denoising effect of these two denoising methods,a group structure dictionary learning denoising method based on graph rule is proposed.The research content of this paper can be summarized as follows:1.In order to solve the problem that it is difficult to remove the noise from the original seismic data,this paper first uses the wavelet transform and Curvelet transform denoising methods in the sparse representation methods based on fixed basis function feature extraction,in which the wavelet transform method has a strong time-frequency analysis ability to effectively represent the local catastrophe features of seismic signals,but this method can not sparsely represent the seismic signals with curvilinear features.The Curvelet transform method has the ability of multi-directional and multiscale representation,so that the in-phase axis of the curve can be well represented while denoising.However,the basis function of the sparse representation method based on fixed basis function is constant and does not have the ability to adapt to the construction basis needed to represent the data.2.In view of the lack of adaptability of traditional fixed basis functions in the process of data sparse representation,this paper uses a dictionary learning method based on data-driven feature learning to sparse seismic data,and mainly studies MOD,KSVD dictionary update algorithm and MP,OMP,LARS sparse solution algorithm.It is found that the vectorization operation of the traditional dictionary learning denoising method for multi-dimensional seismic data will lose the information containing local detail texture features in the seismic data,and the effective signal in the weak signal will be removed as random noise.3.Aiming at the traditional dictionary learning denoising method based on datadriven feature learning,a group structure dictionary learning denoising method based on graph regularization is proposed in this paper.this method limits the similarity between adjacent data by constructing a weight matrix,which ensures that the correlation between atoms is small enough,and the trained breakfast has more correct data structure features and detail features.This method can effectively reconstruct the original data while the seismic data are noisy.In this paper,the effects of denoising and reconstruction are evaluated in twodimensional and three-dimensional theoretical models and actual seismic data.Comparative analysis shows that compared with the traditional sparse representation denoising method,the group structure dictionary learning method based on graph regularity proposed in this paper has better denoising effect.