High-resolution Seismic Imaging Of Complex Media Based On Staining Algorithm And Deep Learning

Seismic migration can achieve high-resolution structural imaging of complex subsurface media,but it will face some difficulties in practical applications.The three main factors affecting seismic imaging are:observed data,migrated velocity models and imaging algorithms.Weak signals are often present in seismic data.These weak signals will cause imaging difficulties.In terms of velocity models,velocity has a large impact on complex media imaging,and accurate velocity helps to obtain high-quality images;conversely,inaccurate velocity leads to distorted images.In terms of migration algorithms,there is often a trade-off between imaging accuracy and computational efficiency,and there are also some imaging artifacts in the migrated images.These problems mentioned above need to be solved urgently.In order to solve the problem of weak signals in seismic data,we study a staining algorithm based on the one-way wave equation.The staining algorithm is a targetoriented wavefields reconstruction,and these reconstructed wavefields are used for target structures imaging.The staining algorithm is used for two-way wave-equation migration,but the two-way wave equation is computationally inefficient and suffers from disturbances of multiples,while the one-way migration does not have these problems.Based on this,we propose a staining algorithm based on the one-way wave equation and use it for weakly illuminated structures imaging.In addition,we also extend the traditional staining wavefield from the source-side to the receiver-side,forming two new staining strategies:receiver-stained strategy and source-receiver staining strategy.Both the proposed two staining strategies can achieve high-resolution imaging of the target structures.To address the problem that the velocity model severely affects the imaging quality,we carry out two main studies.The first study is a detailed investigation of how velocity errors affect imaging quality.In the complex subsurface media,salt structures,for example,are more sensitive to changes in ground stress,and the velocities of sediments around salt bodies are complex.We analyze the velocity changes induced by salt stress perturbations and investigate how these velocity changes affect salt structure imaging.Conventional salt structure imaging methods rarely consider salt-induced stress perturbations and the corresponding velocity perturbations.To demonstrate the importance of salt-induced stress and velocity changes for accurate salt imaging,in this study,we develop a method that couples geomechanical simulations and salt imaging.We first simulate the stress perturbations in the sediments around a salt body using a static geomechanical model,then calculate the corresponding changes in seismic wave velocities based on the stress perturbations,and finally,use reverse time migration for salt imaging and analyze whether the effects of salt stress perturbations and their velocity changes on the imaging were considered.The results show that the velocity variation due to salt stress perturbation has a large impact on salt structure imaging,and therefore,the stress perturbation and velocity variation around the salt should not be neglected during salt structure imaging.The second study is to use deep learning for seismic imaging with velocity errors to address the problem of the large dependence of the migration on the velocity models.Unlike traditional end-to-end networks,we investigate a basis prediction network,which has a distorted common image gathers(CIGs)as input and two basis matrices as output,which performs matrix multiplication to reconstruct pixel-level convolution kernels;applying these convolution kernels to the input distorted CIGs to achieve the CIGs correction,and superimpose the CIGs to obtain the corrected seismic image.Numerical examples demonstrate the effectiveness of the trained network for CIGs correction with velocity errors.The proposed method reduces the sensitivity of migration to the velocity models and obtains high-resolution and accurate images of subsurface structures in the presence of velocity errors.Some research is carried out to address some of the problems inherent in seismic imaging algorithms.In solving the problem of conflicting computational efficiency and imaging accuracy,we use deep learning for enhacing multi-source reverse time migration(RTM).The multi-source RTM methods can solve the problem of RTM computational efficiency,which reduces the computational effort by processing multiple sources simultaneously,but this approach often introduces crosstalk during the imaging process,which seriously interferes with the imaging signals.Plane-wave RTM,as a mainstream multi-source method,reconstructs plane-wave wave fields by performing linear delayed-time on the wavefields and uses these plane-wavefields for migration to obtain plane-wave images of corresponding angles,which will have interference from crosstalk.The superposition of plane wave images from multiple angles can suppress the crosstalk and improve the imaging quality,but it also increases the computational effort,so there is usually a trade-off between computational efficiency and imaging quality when applying plane wave migration.In this study,a convolutional neural network is designed for enhancing plane-wave imaging.The input to the network is 7 plane-wave images at different angles,and the desired output is a clean,high-resolution convolved image,which is obtained by convolving a Ricker wavelet with reflectivities.The proposed method solves the contradiction between computational efficiency and imaging accuracy and achieves high-resolution imaging of complex media at a lower computational cost.For the imaging artifacts in the migration algorithms,we propose a noise-to-noise(N2N)deep learning method for seismic imaging denoising.Conventional deep learning denoising methods achieve image denoising by building a large number of noisyclean image training pairs and then training a regression model,but in many scenes without clean images as labels.The N2N method can solve this problem by composing training sample pairs of noisy-noisy images,and extracting effective signals from these noisy images by self-supervision to remove noise and obtain clean images.For the case of seismic imaging to remove imaging artifacts,clean migrated images are difficult to obtain,and we can only obtain a series of noisy CIGs.Therefore,we investigate the N2N method to compose the noisy CIGs into training sample pairs and achieve artifacts suppression of the noisy CIGs.Numerical and one field examples demonstrate the effectiveness of this self-supervised method in seismic imaging denoising,and different types of imaging artifacts are well removed.