Research On Optimization Of Marchenko Imaging Method And Wavefield Reconstruction

With the continuous advancement and innovation in seismic exploration,seismic imaging technology has increasingly become a core technical means for exploring and developing subsurface resources,especially in the exploration and development of oil and natural gas resources.Seismic imaging technology,through analyzing the propagation of seismic waves in geological media,has revealed the complexity of subsurface structures and significantly improved the precision and efficiency of exploration.In this process,the Green’s function,as a fundamental mathematical tool,is widely applied in the analysis of seismic wavefields,simulating the reflection,refraction,and scattering of waves at different geological interfaces,providing theoretical support for subsurface structure imaging.However,despite decades of significant achievements in seismic imaging technology,traditional methods such as reflection and refraction seismic imaging still face many challenges in dealing with complex stratigraphic structures.These imaging methods often rely on idealized assumptions,such as the homogeneity and layered structure of the medium,when utilizing Green’s functions to simulate seismic wavefields and imaging processes.These assumptions might be valid in simple geological structures,but in complex strata with heterogeneity,anisotropy,and complex surface conditions,they often fail to accurately reflect the actual propagation characteristics of seismic waves.In recent years,the introduction of the Marchenko method has brought new changes to the field of seismic imaging.Based on the concept of autofocusing,this method can effectively extract the information of the upgoing and downgoing wavefields at any point underground from seismic records,thereby achieving fine imaging of subsurface media.A key advantage of the Marchenko method is that the reconstructed Green’s function accurately accounts for multiple scattering effects in complex media,thereby improving the accuracy and resolution of imaging.This makes the Marchenko method demonstrate superior performance compared to traditional imaging methods such as wave equation migration and reverse time migration when dealing with complex media.However,despite the potential demonstrated by the Marchenko method in theoretical research and experimental verification,it still faces challenges and limitations in application.For example,how to effectively reconstruct complete wavefield information from non-uniformly sampled data remains an urgent problem to be solved.The absence of such data not only affects the completeness of the Marchenko wavefield reconstruction but may also lead to inaccurate imaging results,thereby limiting the application potential of this method under complex geological conditions.Additionally,as a computationally intensive technique,the Marchenko method demands high computational resources.Therefore,seeking methods to optimize the Marchenko wavefield reconstruction and imaging process to improve computational efficiency and result accuracy has become a key research direction.In response to the challenges faced by the Marchenko method in application,this thesis employs techniques such as sparse inversion and deep learning to improve the algorithmic structure of Marchenko wavefield reconstruction,thereby enhancing the application efficiency and accuracy of the Marchenko method.Moreover,to strengthen the adaptability of the Marchenko method across various geological environments,this research also explores strategies for transfer learning.The main research outcomes of this thesis are reflected in the following aspects:(1)The traditional Marchenko method,in the process of seismic wavefield reconstruction,usually assumes that the source and receivers are colocated and relies on uniform data sampling.These assumptions are difficult to satisfy in seismic exploration,leading to potential discontinuities and distortions in the focusing functions and Green’s functions under non-ideal sampling conditions,thereby affecting the accuracy of wavefield reconstruction.By introducing the discrete expression of the Marchenko equation,this thesis delves into the sources of error and limitations that incomplete sampling conditions might induce in the standard Marchenko method.Addressing various integration dimensions and subsampling issues,the study proposes a novel sparse constraint Marchenko equation,which is effectively integrated into the Marchenko iterative process.By optimizing the distorted focusing functions and Green’s functions in each iteration,this method successfully achieves Marchenko wavefield reconstruction in the absence of input data,effectively extending the limitations of non-ideal sampling conditions on the application of the Marchenko method.(2)In the Marchenko method,the estimation of direct waves relies on an initial background model.Although this approach is computationally efficient,it often overlooks the diversity of subsurface media complexity,potentially leading to deviations between the estimated results and actual geological structures.These deviations are not limited to the direct waves themselves but may also be further amplified in the computation process of the focusing functions,ultimately affecting the quality of the reconstructed wavefield and imaging results.To address these challenges,this study proposes an enhanced deep learning framework based on feature fusion,the HPAt-net network,to correct errors in the focusing functions and improve the accuracy of Marchenko imaging.The HPAt-net network incorporates a hybrid attention mechanism that effectively processes features of various scales and retains key information by fusing different attention modules.Additionally,the framework integrates a Haar Wavelet Downsampling(HWD)module and a Pixel Attention Guided Fusion(PAG)module,which not only enhance the model’s feature extraction capabilities but also facilitate effective information fusion at different levels,thereby capturing the details of geological structures.Experimental results demonstrate that the HPAt-net network can effectively identify and correct errors in the focusing functions,thereby improving the precision and reliability of imaging results.(3)One of the main challenges faced by deep learning methods in seismic data processing,especially in the detection of complex geological structures,is the scarcity of labeled data.Given the complexity of subsurface structures and the difficulties in collecting seismic data,the data obtained are often limited in quantity,which restricts the effectiveness of directly applying deep learning techniques to the Marchenko method.Transfer learning technology offers an effective solution to this problem.By transferring existing knowledge and experience to new tasks,transfer learning can reduce the dependence on a large amount of labeled data,thus achieving accurate geological imaging even in situations with limited data.This thesisdelves into how to use transfer learning to optimize the performance of the Marchenko method in multiscenario applications.In key processes such as focusing function correction and travel time error handling,the accuracy and efficiency of the Marchenko method are enhanced by selecting suitable source tasks,adjusting transfer strategies,and fine-tuning model parameters.Moreover,since the precise labels required by the Marchenko method are difficult to obtain and can only rely on results from traditional seismic exploration methods as label data,which typically contain certain degrees of uncertainty and errors and cannot fully accurately reflect the actual situation of subsurface structures.Directly using these imprecise labels for training deep learning models may lead to limited accuracy in the imaging results.Addressing this issue,this study also explores the potential application of transfer learning in dealing with imprecise label data.By appropriately fine-tuning based on pre-trained models,the model can be better adapted to the specific feature distribution and data structure of the domain,thereby achieving more accurate Marchenko imaging under the guidance of imprecise labels.In summary,this thesis enhances the accuracy and efficiency of Marchenko wavefield reconstruction and imaging under non-ideal sampling conditions and complex geological structures through improvements to the traditional Marchenko method,the introduction of deep learning frameworks,and the application of transfer learning techniques.