Deep Learning-based Seismic Data Processing Techniques And Imaging Research

Seismic exploration,as an important field of geophysics,aims to re-veal the physical parameters distribution of subsurface media through the collec-tion and analysis of seismic data,achieving high-resolution imaging and indepth interpretation of underground structures.This technology plays a crucial role in the exploration and development of resources such as petroleum,natural gas,coal,and minerals,which are essential for the prosperity and development of society and economy.In order to further enhance the accuracy and efficiency of seismic exploration,high-resolution seismic imaging algorithms have emerged.By utiliz-ing advanced seismic data processing techniques,rich information in seismic data can be fully exploited,and the resolution and signal-to-noise ratio of seismic ex-ploration can be improved,enabling the restoration of more accurate underground structural information and achieving higher precision imaging of the subsurface.This article proposes a deep learning network framework-driven processing and imaging algorithm for seismic data pre-processing and high-resolution imag-ing.Detailed research and analysis are conducted from three aspects: seismic data pre-processing,seismic reverse time migration imaging,and full waveform inversion imaging.The high-resolution imaging technology for seismic profile pre-processing is primarily discussed in this article.Firstly,a deep learning framework for seismic profile denoising is established and optimized.The effectiveness of this frame-work is demonstrated by comparing it with traditional mainstream algorithms.Secondly,a deep learning framework for the separation of direct waves in seis-mic profiles is constructed,laying the foundation for subsequent research.Sub-sequently,a corresponding deep learning framework is established to address the issue of missing traces that may occur during seismic acquisition,and its spe-cific process and operational details are elaborated in detail.The superiority of this framework is demonstrated by comparing it with traditional mainstream algorithms.Finally,the deep learning framework is used to achieve super-resolution reconstruction of seismic profiles,which improves the resolution of the original profile,avoids the problem of discontinuity caused by the lack of traces,and lays the foundation for high-resolution seismic imaging.By implementing the above preprocessing process,this article achieves pre-processing imaging of seismic profiles,providing strong support for subsequent work.To address the issues of interference and imaging efficiency in traditional reverse time migration imaging algorithms in seismic profile processing,this ar-ticle proposes an improved scheme and conducts reverse time migration process-ing.Firstly,the problem of interference in traditional reverse time migration algo-rithms is improved using methods such as Helmholtz decomposition and Poynting vector decomposition.Subsequently,the pre-processed seismic profiles are used for reverse time migration imaging,and the new algorithm is shown to improve imaging quality compared with traditional algorithms.To address the issue of memory consumption in reverse time migration,a wave field data compression strategy based on compressed sensing is proposed,effectively reducing physi-cal memory consumption.In addition,to further improve the imaging quality of reverse time migration,an enhanced post-processing deep learning framework is proposed,achieving further improvement in the resolution of seismic profile reverse time migration imaging.Through the above processing processes,high-precision and high-efficiency reverse time migration imaging of seismic profiles is achieved.Based on pre-processed seismic profiles,full waveform inversion imaging is carried out,and traditional algorithms are improved to address the issues of low-frequency loss and imaging efficiency in full waveform inversion algorithms.Firstly,to address the computational efficiency issue of full waveform inversion,a strategy of random source and source encoding full waveform inversion is pro-posed to improve computational efficiency.Secondly,to address the issue of low-frequency data loss in full waveform inversion,a deep learning framework is applied to propose a low-frequency data prediction strategy.Subsequently,to address the low efficiency issue of full waveform inversion,a full waveform inversion framework based on time-space multi-scale is proposed,making full use of key information in data and achieving a significant improvement in ef-ficiency.Finally,to address the ill-posed problem of traditional algorithms,a resolution-enhancing regularization full waveform inversion strategy is proposed,which greatly improves imaging resolution.Through these processing processes,a high-precision and high-efficiency full waveform inversion imaging algorithm flow for seismic profiles is established.Overall,this article proposes a series of techniques and methods in the field of seismic data pre-processing and high-resolution imaging,which plays a crucial role in improving the imaging resolution and efficiency of seismic exploration.This article not only contributes novel technical means to the fields of geophysics and seismology but also provides strong support for resource exploration and de-velopment in relevant industries.Additionally,the methods and algorithms in-volved in this article provide new ideas and inspiration for imaging problems in other fields,with broad potential applications.There are 183 figures,23 tables,and 221 citations in this thesis.