Separation Of Marine Multi-Source Blended Wavefields Based On Iterative Shaping Regularization

Compared with marine single-source seismic acquisition technology,marine multisource seismic acquisition technology has attracted increasing attention from major oil and gas companies because of its advantages in reducing exploration time,improving data quality,increasing the flexibility of surveys,and so on.However,there is wavefield interference in the seismic data obtained by multi-source acquisition,which is difficult to process directly by traditional processing methods effectively.Currently,the primary way to deal with multi-source blended data is to separate it into the form of single-source and then use the conventional processing flow for subsequent imaging processing.Therefore,how to effectively separate the multi-source blended wavefields is the key to the success of this method.To accelerate the practical application of marine multi-source acquisition technology,this thesis has studied the separation methods of marine multi-source blended wavefields.As the current marine multi-source acquisition typically uses random time delays to encode sources,it results in continuous records of the primary source and approximately random distribution of other sources in non-common shot gathers.This thesis is based on this property to separate blended wavefields.At the same time,this thesis regards the separation problem as an inversion problem,but for marine multi-source blended data,the inverse problem is usually underdetermined,and constraints need to be added.This thesis uses an iterative framework based on shaping regularization to solve the inverse separation problem.In this iterative framework,three different shaping operators are constructed according to the sparsity,nonlocal self-similarity,and low-rank property of seismic data,resulting in three iterative inversion separation algorithms with different separation effects:The shaping operator in the first iterative inversion separation algorithm consists of a thresholding operator in the contourlet transform domain.The algorithm is based on the sparsity prior to seismic data: the transform coefficients corresponding to the signal of the primary source and the blending noise of other sources have different characteristics,so the thresholding operator can be used to shrink the transform coefficients to remove the coefficients corresponding to the blending noise and achieve the purpose of blended wavefield separation.This method’s thresholding operator comprises a Bayesian threshold estimation method with scale,direction,and spatial adaptability and a new thresholding function.The proposed adaptive Bayesian threshold estimation method fully considers the distribution characteristics of blended data in the contour wave transform domain and can obtain more accurate thresholds.At the same time,the newly constructed thresholding function is flexible between the soft and hard thresholding functions.It can shrink the transform coefficients more finely according to the obtained thresholds.The test results show that this method has high computational efficiency and can preserve more signals while removing blending noise.The shaping operator in the second iterative inversion separation algorithm consists of the matching grouping and the thresholding operator in the 3D sparse transform domain.The algorithm is based on the nonlocal self-similarity and sparsity prior to the seismic data: firstly,the seismic data is divided into blocks,and then the 3D similar groups composed of similar blocks are obtained by block matching according to the nonlocal self-similarity of the seismic data.For the obtained similar groups,the corresponding 3D sparse transform is constructed to represent it sparsely,and the corresponding thresholding operator is built to shrink the 3D transform coefficients.Compared with the2 D sparse transform,which only uses the local correlation of seismic data,the 3D sparse transform can use the local correlation and nonlocal self-similarity of seismic data at the same time.The obtained transform coefficients have higher sparsity and larger amplitude values,making the thresholding operator more effective for processing 3D transform coefficients.The test results show that the algorithm can effectively separate blended wavefields and has a good recovery ability for weak signals masked by strong blending noise.The shaping operator in the third iterative inversion separation algorithm consists of matching grouping,nuclear norm minimization,and compensating for the loss of signals.The algorithm is based on the nonlocal self-similarity and low-rank prior to seismic data:the nonlocal self-similarity of seismic data can be reflected in higher sparsity and characterized by the low-rank property,and low-rank similar group matrices can be constructed by vectorizing the data blocks in similar groups.When blending noises from other sources exist,the low-rank structures of similar group matrices are destroyed,and the rank of the matrix is increased.Therefore,the separation problem is transformed into a rank minimization problem and further into a weighted nuclear norm minimization problem,which is solved by the singular value thresholding algorithm.Then,the local similarity between the deblended data and the blending noise removed is calculated to judge the leakage of the signals,and a corresponding weighting operator is constructed to extract the lost signals from the blending noise removed and return them to the deblended data,which further improves the quality of the deblended data.The test results show that the algorithm has a better separation effect and can recover more signals.The novelties of this thesis are summarized as follows:(1)Improved the traditional Bayesian threshold estimation method to have scale,direction,and spatial adaptability,allowing it to adaptively adjust the corresponding thresholds based on the distribution characteristics of blended data in the contourlet transform domain;proposed a new thresholding function that varies flexibly between soft and hard thresholding functions,allowing for more precise threshold shrinkage processing of the transform coefficients based on the estimated thresholds.(2)3D similar groups are extracted from blended data according to nonlocal selfsimilarity,and a 3D sparse transform is constructed to represent the similar groups sparsely so as to obtain a higher sparsity than the 2D sparse transform.A corresponding thresholding operator was also constructed to shrink the coefficients in the 3D transform domain.The higher sparsity and larger amplitude of the 3D transform coefficients make the thresholding operator more effective at shrinking the coefficients.(3)Extracting similar group matrices with low-rank structure from blended data according to nonlocal self-similarity,converting the multi-source blended wavefield separation problem into the rank minimization problem,which is further converted into the weighted nuclear norm minimization problem,and solving it by the singular value thresholding algorithm.Then extracting the lost signals from the removed blending noise and returning them to the deblended data further improves the quality of the deblended data.