Research Of Reflectivity Inversion With Automatic Spatially Correlated Evaluation Based On Sparse Bayesian Learning

With the development of new techniques and methods in geophysics exploration area,the conventional seismic reflection exploration always plays a dominant role and related researches about seismic data processing and interpretation does not stagnate in certain time.The elaborate seismic data and detailed image corresponding to subsurface layers in complex geology region can guarantee the achievement of prospection investigation goal.As the essential procedure of seismic data processing,high resolution processing technique could eliminate most interference and provide the precise seismic poststack dataset or cube.Its advantage and deployment accuracy determines the quality of output data directly.Therefore,the identification of thin layer and weak reflector from final record is the overcome difficulty in the destination of oil-gas field exploration.The requirement of stable and efficient approach still promotes the acquisition of reflection information of underground interfaces.The poststack seismic reflection signal is expressed as a superposition calculation between spike response of reflection interfaces and seismic waveform.This process can be described as a band-pass filtering.As the significant procedure to eliminate the filtering effect,reflectivity inversion or deconvolution could indicate and estimate the position and elastic parameter value of reflector.However,the band-limited of seismic wavelet and absorption attenuation of layer lead to the ill-condition and non-uniqueness of this inversion problem.Thus,that how to utilize the reasonable assumptions of geology model and to research the excellent algorithms should be emphasized in the whole course of seismic exploration.This thesis focuses on the multichannel reflectivity inversion and deconvolution for poststack seismic data and proposes two inversion approaches for stationary and attenuated data.The first approach is the multichannel spatially correlated reflectivity inversion using block sparse Bayesian learning(bSBL)and the other is multichannel block sparse Bayesian learning reflectivity inversion with lp norm(0<p<1)criterion-based Q estimation.The reflectivity inversion approach based on a variety of regularization terms was extensively developed and applied to image subsurface structure over the recent years.In addition,multichannel reflectivity inversion or deconvolution considering the lateral continuity of reflection interfaces or reflectivity in adjacent channels has been developed.However,these processing operations seldom adaptively judge the stratal continuity or automatically alter the parameters of the corresponding algorithm.To use the special correlation of the reflection information contained in the seismic data,a multichannel spatially correlated reflectivity inversion using block sparse Bayesian learning(bSBL)is introduced in the initial part of this thesis.The method adopts a covariance matrix that describes the spatial relationship of reflectivity and simultaneously controls the temporal sparsity.With an expectation maximization(EM)algorithm,we can obtain the parameters of the multichannel reflectivity model,including the mean(i.e.,the estimated multichannel reflectivity)and the covariance matrix(i.e.,the correlation of nonzero reflection impulses).The noise variance in the observed seismic data is also estimated during the inversion processing.Due to the contribution of reflectivity correlation in different traces,the performance of the multichannel spatially correlated reflectivity inversion using bSBL is significantly superior to the trace-by-trace processing method in the presence of a medium level of noise.The synthetic and real data examples illustrate that the lateral continuity is well preserved in seismic profiles after inversion.The next part of this thesis is the illustration of multichannel reflectivity inversion for the attenuated seismic reflection data which can be described via a nonstationary convolution model with quality factor Q.In according with the linear matrix-matrix multiplication operation that is deduced to depict the multichannel nonstationary seismic received signal,we present a spatially correlated reflectivity inversion with Q estimation based on block sparse Bayesian learning(bSBL)and lp norm(0<p<1)criterion.In contrast to pre-existing time-variant deconvolution,the proposed technique can eliminate the wavelet-filtering and Q-filtering effect simultaneously and retrieve an optimal reflectivity matrix without providing Q value by anticipation.Through building the lp norm(0<p<1)criterion and scanning Q strategy,we could capture the optimal Q to calculate the blurring operator in inversion function.The inverted reflectivity result will be satisfied with presupposition that multi-trace reflectivity series is comparatively sparse corresponding to its minimum lp norm(0<p<1)when the accurate Q is applied to build attenuation formula.In reflectivity inversion area,the relationship among reflection spikes in adjacent traces as a prior information is represented by covariance matrix of reflectivity model in bSBL framework and assists in solving procedure to promote the inversion precision.To diminish the influence of man-made parameter selection,the hyperparameters of reflectivity and noise model are estimated in virtue of EM algorithm.The contribution of spatial correlation could guarantee a higher quality of reflectivity image.New method merges the reflectivity inversion and Q estimation into a single processing and avoids the drawback of conventional Q extraction technique.Synthetic and field data sets also prove the practicality of developed technique and indicate the favorable anti-noise capability.On the other hand,the lp norm constraint is utilized to assist the multichannel nonstationary seismic data acoustic impedance(AI)inversion.The application area of lp norm is extended with the new approach.