Jiaxing Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Jiaxing Sun has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Jidong Yang
State Key Laboratory of Deep Oil and Gas, School of Geosciences, China University of Petroleum (East China), Qingdao, China
9 shared publications - Jianping Huang
National Key Laboratory of Deep Oil and Gas, School of Geosciences, China University of Petroleum (East China), Qingdao, Shandong, China
7 shared publications - Youcai Yu
School of Geosciences, the Laboratory for Marine Mineral Resources, and the National Laboratory for Marine Science and Technology, China University of Petroleum, Qingdao, China
3 shared publications Chong Zhao
National Key Laboratory of Deep Oil and Gas, China University of Petroleum, Qingdao, China
2 shared publications- Shanyuan Qin
State Key Laboratory of Deep Oil and Gas, School of Geosciences, China University of Petroleum (East China), Qingdao, China
2 shared publications Xuanhao Chen
National Key Laboratory of Deep Oil and Gas, China University of Petroleum, Qingdao, China
2 shared publications
Latest Publications
- 15 June 2026
- Published by Wiley in Geophysical Prospecting
- Vol. 74 (5) , e70205
Deep‐Learning‐Based Migration Velocity Analysis With Multi‐Type Input Data
In seismic data processing, velocity model building is always at its core and its accuracy directly affects the quality of subsequent imaging and reservoir characterization. Conventional ray‐ and wave‐equation‐based migration velocity analysis (MVA) methods that aim to flatten angle‐domain common‐image gathers (ADCIGs) require repeated application of prestack depth migration, resulting in a significant computational challenge. To mitigate this issue, we present a deep‐learning‐based MVA approach to efficiently generate accurate velocity models. Instead of relying solely on the ADCIGs as in conventional MVA, we use multi‐type data, including the original ADCIGs, migration velocity model, the envelope functions of the Radon transform of ADCIGs and RGT models, as network inputs, which can provide diverse information to effectively enhance the accuracy of velocity inversion. In addition, to avoid unphysical lateral discontinuities, the multi‐type data of several horizontally neighbouring locations rather than a single location are utilized during the network training and prediction. Numerical experiments show that the proposed deep‐learning‐based MVA method can not only recover the low‐ and middle‐wavenumber velocity model as in conventional MVA approaches but also has the potential to generate high‐wavenumber information at the strata interfaces. This improvement can be attributed to the strong nonlinear mapping capability of deep learning, which effectively relates the wide‐spectrum velocity information in the label data to ADCIG‐related multi‐type information in the training data. The high‐quality migration results calculated with the velocity models predicted by the network verify the feasibility of the proposed method.
- 15 June 2026
- Published by Wiley in Geophysical Prospecting
- Vol. 74 (5) , e70205
Deep‐Learning‐Based Migration Velocity Analysis With Multi‐Type Input Data
In seismic data processing, velocity model building is always at its core and its accuracy directly affects the quality of subsequent imaging and reservoir characterization. Conventional ray‐ and wave‐equation‐based migration velocity analysis (MVA) methods that aim to flatten angle‐domain common‐image gathers (ADCIGs) require repeated application of prestack depth migration, resulting in a significant computational challenge. To mitigate this issue, we present a deep‐learning‐based MVA approach to efficiently generate accurate velocity models. Instead of relying solely on the ADCIGs as in conventional MVA, we use multi‐type data, including the original ADCIGs, migration velocity model, the envelope functions of the Radon transform of ADCIGs and RGT models, as network inputs, which can provide diverse information to effectively enhance the accuracy of velocity inversion. In addition, to avoid unphysical lateral discontinuities, the multi‐type data of several horizontally neighbouring locations rather than a single location are utilized during the network training and prediction. Numerical experiments show that the proposed deep‐learning‐based MVA method can not only recover the low‐ and middle‐wavenumber velocity model as in conventional MVA approaches but also has the potential to generate high‐wavenumber information at the strata interfaces. This improvement can be attributed to the strong nonlinear mapping capability of deep learning, which effectively relates the wide‐spectrum velocity information in the label data to ADCIG‐related multi‐type information in the training data. The high‐quality migration results calculated with the velocity models predicted by the network verify the feasibility of the proposed method.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64, 5922415
STLSM: A Swin Transformer-Based Framework for Least-Squares Seismic Imaging
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64, 5922415
STLSM: A Swin Transformer-Based Framework for Least-Squares Seismic Imaging
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-11
Elastic Wavefield Decomposition Using the Physical-Constrained Neural Network and Its Application on Reverse-Time Migration
By using multicomponent and multiwave seismic data, elastic reverse time migration (ERTM) can produce accurate PP and PS images for complex structures. One key of ERTM is to decompose the source- and receiver-side wavefields into pure P- and S-waves with correct amplitudes and phases. Conventional wavefield decomposition methods require to solve either a Poisson’s equation or an additional P-wave equation, increasing the computational cost. To mitigate this issue, we present a novel framework of physical constrained neural network (PCNN) to efficiently implement P- and S-wave separation. The new network includes both fully connected and convolutional blocks for data fitting and takes into account the properties of curl-free P-wave and divergence-free S-wave as a physical constraint. After the training, the proposed PCNN can accurately decompose elastic wavefield and produce pure P- and S-waves with correct phases, amplitudes and physical unit. We then apply the PCNN-based method to multicomponent reverse-time migration, in which a modified dot-product imaging condition is used to calculate PP and PS images. Numerical experiments demonstrate that the proposed PCNN-based workflow can produce accurate PP and PS images while improving the computational efficiency by three orders of magnitude compared to traditional Helmholtz-based decomposition methods.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-11
Elastic Wavefield Decomposition Using the Physical-Constrained Neural Network and Its Application on Reverse-Time Migration
By using multicomponent and multiwave seismic data, elastic reverse time migration (ERTM) can produce accurate PP and PS images for complex structures. One key of ERTM is to decompose the source- and receiver-side wavefields into pure P- and S-waves with correct amplitudes and phases. Conventional wavefield decomposition methods require to solve either a Poisson’s equation or an additional P-wave equation, increasing the computational cost. To mitigate this issue, we present a novel framework of physical constrained neural network (PCNN) to efficiently implement P- and S-wave separation. The new network includes both fully connected and convolutional blocks for data fitting and takes into account the properties of curl-free P-wave and divergence-free S-wave as a physical constraint. After the training, the proposed PCNN can accurately decompose elastic wavefield and produce pure P- and S-waves with correct phases, amplitudes and physical unit. We then apply the PCNN-based method to multicomponent reverse-time migration, in which a modified dot-product imaging condition is used to calculate PP and PS images. Numerical experiments demonstrate that the proposed PCNN-based workflow can produce accurate PP and PS images while improving the computational efficiency by three orders of magnitude compared to traditional Helmholtz-based decomposition methods.
- 11 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-8
Acoustic and Elastic Reverse-Time Migration With an Angle-Related Imaging Condition for Imaging Steeply-Dipping Structures
The subsurface steeply-dipping structures, such as buried hills, fault zones, and salt flanks, are difficult to accurately image in seismic exploration due to weak illumination from large incident angles. To mitigate this issue, we propose an angle-related imaging condition for acoustic and elastic reverse time migration (RTM) to improve the image quality of steeply-dipping structures. We first use the wavefield decomposition method based on the Hilbert transform to decompose the extrapolated wavefield into up-down-left-right going components. We observed that the left-and right-going wavefields can accurately image steeply-dipping structures, while the up- and down-going wavefields mainly contribute to the layers with small dipping angles. Then, an angle-related imaging condition is developed to use the decomposed directional wavefields to enhance the image quality of both flat layers and steeply-dipping structures. In elastic reverse time migration (ERTM), after the directional wavefield decomposition, the vector Helmholtz decomposition is used to decouple PS waves to produce PP and PS images. Numerical examples demonstrate that the proposed method produces more accurate images than conventional RTM for steeply-dipping structures.
- 11 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-8
Acoustic and Elastic Reverse-Time Migration With an Angle-Related Imaging Condition for Imaging Steeply-Dipping Structures
The subsurface steeply-dipping structures, such as buried hills, fault zones, and salt flanks, are difficult to accurately image in seismic exploration due to weak illumination from large incident angles. To mitigate this issue, we propose an angle-related imaging condition for acoustic and elastic reverse time migration (RTM) to improve the image quality of steeply-dipping structures. We first use the wavefield decomposition method based on the Hilbert transform to decompose the extrapolated wavefield into up-down-left-right going components. We observed that the left-and right-going wavefields can accurately image steeply-dipping structures, while the up- and down-going wavefields mainly contribute to the layers with small dipping angles. Then, an angle-related imaging condition is developed to use the decomposed directional wavefields to enhance the image quality of both flat layers and steeply-dipping structures. In elastic reverse time migration (ERTM), after the directional wavefield decomposition, the vector Helmholtz decomposition is used to decouple PS waves to produce PP and PS images. Numerical examples demonstrate that the proposed method produces more accurate images than conventional RTM for steeply-dipping structures.
- 5 April 2024
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 89 (3) , S235-S250
Generative adversarial network-enhanced directional seismic wavefield decomposition and its application in reverse time migration
Reverse time migration (RTM) is an accurate method for imaging complex geologic structures without imposing any dip limitations. However, a large amount of high-amplitude low-frequency noise, which is mainly generated by the crosscorrelation of source and receiver wavefields propagating in the same directions, seriously contaminates the image quality. The causal imaging condition with separated up- and downgoing wavefields is an effective approach to reduce these low-frequency artifacts. Explicit up- and downgoing wavefield decomposition based on the Hilbert transform is computationally expensive due to additional wavefield extrapolation and storage for the imaginary parts. Directionally propagating wavefields have distinctive kinematic patterns such as traveltime and wavefront curvature, which provide us an opportunity to implement the wavefield decomposition using the statistical neural network method. Using extrapolated wavefields as the input and the decomposed up-, down-, left-, and rightgoing wavefields as the labeled data, we train a pair of generative adversarial networks to predict the directional wavefields. The training data sets are generated using seismic full-waveform modeling and explicit wavefield decomposition based on the Hilbert transform. Then, the decomposed directional wavefields are incorporated into a novel imaging condition that depends on the subsurface dip angles to compute the reflectivity perpendicular to the reflectors. Numerical experiments demonstrate that our method can produce accurate directional wavefield decomposition results and high-quality reflectivity images without low-wavenumber artifacts.
- 5 April 2024
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 89 (3) , S235-S250
Generative adversarial network-enhanced directional seismic wavefield decomposition and its application in reverse time migration
Reverse time migration (RTM) is an accurate method for imaging complex geologic structures without imposing any dip limitations. However, a large amount of high-amplitude low-frequency noise, which is mainly generated by the crosscorrelation of source and receiver wavefields propagating in the same directions, seriously contaminates the image quality. The causal imaging condition with separated up- and downgoing wavefields is an effective approach to reduce these low-frequency artifacts. Explicit up- and downgoing wavefield decomposition based on the Hilbert transform is computationally expensive due to additional wavefield extrapolation and storage for the imaginary parts. Directionally propagating wavefields have distinctive kinematic patterns such as traveltime and wavefront curvature, which provide us an opportunity to implement the wavefield decomposition using the statistical neural network method. Using extrapolated wavefields as the input and the decomposed up-, down-, left-, and rightgoing wavefields as the labeled data, we train a pair of generative adversarial networks to predict the directional wavefields. The training data sets are generated using seismic full-waveform modeling and explicit wavefield decomposition based on the Hilbert transform. Then, the decomposed directional wavefields are incorporated into a novel imaging condition that depends on the subsurface dip angles to compute the reflectivity perpendicular to the reflectors. Numerical experiments demonstrate that our method can produce accurate directional wavefield decomposition results and high-quality reflectivity images without low-wavenumber artifacts.
- 6 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-9
PsfDeconNet: High-Resolution Seismic Imaging Using Point-Spread Function Deconvolution With Generative Adversarial Networks
Least-squares migration (LSM) aims to seek the best-fit solution for subsurface reflectivity with high image resolution and balanced amplitudes by minimizing the mismatching between synthetic and observed seismic data. It can be implemented either in data domain or in image domain. Data-domain LSM iteratively updates reflectivity using the gradient-based algorithms. However, it requires expensive computation cost to converge to a good solution, which is still challenging for large-scale datasets under current computational capacity. Point spread function (PSF) deconvolution is an efficient and accurate image-domain LSM approach to reduce migration artifacts caused by the limited aperture and the finite frequency wavelet and improve image resolution. But seismic velocity models have millions of grid points, which makes it prohibitively expensive to directly compute the PSF and its deconvolution operator. To obtain high-resolution images and reduce computing cost, we develop a deep-learning-based method to directly approximate the deconvolution operator of each separated PSF. First, we calculate migrated image and PSFs using one pass of seismic modeling and traditional adjoint migration. Next, we train a conditional generative adversarial networks (cGANs) by using the migrated images, PSFs and migration velocity as the input and the true reflectivity as the labeled data. At last, the trained network is applied to the migrated images, PSFs and migration velocity of test datasets to predict the reflectivity model. With the well-trained cGANs, we can predict high-quality LSM images efficiently and save considerable computational cost. Numerical examples for synthetic models and field data demonstrate that the proposed method can accurately predict PSF deconvolution operators and provide high-quality deblurred LSM images with significantly reduced computational and memory costs.
- 6 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-9
PsfDeconNet: High-Resolution Seismic Imaging Using Point-Spread Function Deconvolution With Generative Adversarial Networks
Least-squares migration (LSM) aims to seek the best-fit solution for subsurface reflectivity with high image resolution and balanced amplitudes by minimizing the mismatching between synthetic and observed seismic data. It can be implemented either in data domain or in image domain. Data-domain LSM iteratively updates reflectivity using the gradient-based algorithms. However, it requires expensive computation cost to converge to a good solution, which is still challenging for large-scale datasets under current computational capacity. Point spread function (PSF) deconvolution is an efficient and accurate image-domain LSM approach to reduce migration artifacts caused by the limited aperture and the finite frequency wavelet and improve image resolution. But seismic velocity models have millions of grid points, which makes it prohibitively expensive to directly compute the PSF and its deconvolution operator. To obtain high-resolution images and reduce computing cost, we develop a deep-learning-based method to directly approximate the deconvolution operator of each separated PSF. First, we calculate migrated image and PSFs using one pass of seismic modeling and traditional adjoint migration. Next, we train a conditional generative adversarial networks (cGANs) by using the migrated images, PSFs and migration velocity as the input and the true reflectivity as the labeled data. At last, the trained network is applied to the migrated images, PSFs and migration velocity of test datasets to predict the reflectivity model. With the well-trained cGANs, we can predict high-quality LSM images efficiently and save considerable computational cost. Numerical examples for synthetic models and field data demonstrate that the proposed method can accurately predict PSF deconvolution operators and provide high-quality deblurred LSM images with significantly reduced computational and memory costs.
- 9 January 2024
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 89 (5) , S379-S390
Image-domain least-squares imaging in an elastic vertically transversely isotropic medium: Multiparameter point-spread function deconvolution
By solving a linear inverse problem, least-squares migration (LSM) can provide higher-precision images of complex subsurface structures than traditional adjoint-operator-based migration. LSM can be conducted in the data domain and the image domain (IDLSM). IDLSM in acoustic media has been extensively studied, commonly using single-parameter point-spread function (PSF) deconvolution. For high-accuracy imaging of complex media, we develop an image-domain elastic least-squares imaging method using multiparameter PSF deconvolution for vertically transversely isotropic (VTI) media. The multiparameter PSFs, including P-P, P-S, S-P, and S-S components, are calculated on a sparse grid using elastic VTI wave equation Born modeling and migration. The traditional migration result is then decomposed into individual local results, each corresponding to the size of a single PSF, through unit partitioning. Finally, a set of filters is derived from four elastic VTI PSF components to approximate the Hessian inverse, which are subsequently assembled to reconstruct an entire deconvolved image. The numerical experiments confirm that our method can significantly decrease multiparameter crosstalk artifacts and improve spatial resolution and amplitude fidelity.
- 9 January 2024
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 89 (5) , S379-S390
Image-domain least-squares imaging in an elastic vertically transversely isotropic medium: Multiparameter point-spread function deconvolution
By solving a linear inverse problem, least-squares migration (LSM) can provide higher-precision images of complex subsurface structures than traditional adjoint-operator-based migration. LSM can be conducted in the data domain and the image domain (IDLSM). IDLSM in acoustic media has been extensively studied, commonly using single-parameter point-spread function (PSF) deconvolution. For high-accuracy imaging of complex media, we develop an image-domain elastic least-squares imaging method using multiparameter PSF deconvolution for vertically transversely isotropic (VTI) media. The multiparameter PSFs, including P-P, P-S, S-P, and S-S components, are calculated on a sparse grid using elastic VTI wave equation Born modeling and migration. The traditional migration result is then decomposed into individual local results, each corresponding to the size of a single PSF, through unit partitioning. Finally, a set of filters is derived from four elastic VTI PSF components to approximate the Hessian inverse, which are subsequently assembled to reconstruct an entire deconvolved image. The numerical experiments confirm that our method can significantly decrease multiparameter crosstalk artifacts and improve spatial resolution and amplitude fidelity.
- 10 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 20 (1545598X) , 1-5
Multitask Deep Learning for Least-Squares Imaging: Seismic Reflectivity Inversion and Quantitative Error Analysis
Least-squares migration (LSM) produces an accurate solution for subsurface reflectivity by solving a linear inverse problem, which can effectively mitigate the unbalanced subsurface illumination, suppress acquisition footprint and improve image resolution. It can be implemented in the data or image domain. Data-domain LSM requires intensive computations for many times of Born modeling and adjoint migration. By contrast, image-domain LSM requires a large memory storage for calculating and saving the Hessian matrix. Due to the ability of describing sophisticated nonlinear relations, the neural network can be used to approximate Hessian inverse by using the reverse time migration (RTM) image as the input and the true reflectivity model as the output. However, previous neural network-based LSMs mainly focused on the enhancement of image resolution and amplitude fidelity, but few on error and uncertainty analysis of network predicted images. We present a novel multi-task deep learning method for the image-domain LSM to simultaneously estimate Hessian inverse and evaluate kinematic and dynamic errors of network predicted results. We first use the RTM images along with migration velocity and source illumination as the network input and regard both the true reflectivity model and kinematic or dynamic error between predicted and labeled images as the output to train the network weights. Then, the L2-norm misfits of two terms are incorporated in the loss function to calculate the gradient to optimize the network trainable weights for producing an accurate LSM image. Numerical examples demonstrate that as the epoch of network training increases, the imaging accuracy of subsurface reflectors is gradually enhanced and the kinematic and dynamic errors of network predicted results are reduced, suggesting that the proposed multi-task deep learning method is effective in simultaneously predicting high-quality LSM image and conducting quantitative error analysis.
- 10 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 20 (1545598X) , 1-5
Multitask Deep Learning for Least-Squares Imaging: Seismic Reflectivity Inversion and Quantitative Error Analysis
Least-squares migration (LSM) produces an accurate solution for subsurface reflectivity by solving a linear inverse problem, which can effectively mitigate the unbalanced subsurface illumination, suppress acquisition footprint and improve image resolution. It can be implemented in the data or image domain. Data-domain LSM requires intensive computations for many times of Born modeling and adjoint migration. By contrast, image-domain LSM requires a large memory storage for calculating and saving the Hessian matrix. Due to the ability of describing sophisticated nonlinear relations, the neural network can be used to approximate Hessian inverse by using the reverse time migration (RTM) image as the input and the true reflectivity model as the output. However, previous neural network-based LSMs mainly focused on the enhancement of image resolution and amplitude fidelity, but few on error and uncertainty analysis of network predicted images. We present a novel multi-task deep learning method for the image-domain LSM to simultaneously estimate Hessian inverse and evaluate kinematic and dynamic errors of network predicted results. We first use the RTM images along with migration velocity and source illumination as the network input and regard both the true reflectivity model and kinematic or dynamic error between predicted and labeled images as the output to train the network weights. Then, the L2-norm misfits of two terms are incorporated in the loss function to calculate the gradient to optimize the network trainable weights for producing an accurate LSM image. Numerical examples demonstrate that as the epoch of network training increases, the imaging accuracy of subsurface reflectors is gradually enhanced and the kinematic and dynamic errors of network predicted results are reduced, suggesting that the proposed multi-task deep learning method is effective in simultaneously predicting high-quality LSM image and conducting quantitative error analysis.
- 14 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 61 (01962892) , 1-13
LsmGANs: Image-Domain Least-Squares Migration Using a New Framework of Generative Adversarial Networks
Compared with traditional adjoint migration, the least-squares migration (LSM) can effectively mitigate the unbalanced illumination and limited resolution associated with finite acquisition apertures, complex overburden structures and band-limited records. Data-domain LSM needs many times of Born modeling and adjoint migration to converge to a good solution, which is still challenging for large-scale 3D model under current computational capacity. To reduce computational cost and produce high-quality images, we directly approximate the Hessian inverse in the image-domain LSM using a new framework of generative adversarial networks (GANs). The migrated images, source illumination and migration velocity model are used as input data for the GANs, and the ground-truth reflectivity is utilized as the label data to train the network. Directly applying conventional GAN framework to implement the image-domain LSM leads to dislocated reflection events and incorrect images. To overcome this issue, we develop a new GAN framework that is more suitable for the Hessian approximation of image-domain LSM, which is named as LsmGANs. In the new framework, we use a max-pooling instead of convolution to downsample the feature maps to capture horizontal and vertical variations of reflectors. This enables us to map reflection events to correct location in downsampling. To address the lateral discontinuity of events in the predicted image from conventional GANs, we further apply multiple transform layers to strengthen feature transformation to guide Hessian approximation. Finally, we add the skip connection in the transform layer to enhance the information exchange of the feature channels and avoid the gradient vanishing problem to improve image resolution. Assembling predicted patches to construct a whole reflectivity image is a key step in the neural-network-based LSM. We investigate four strategies using different overlapping ratio and window functions to assemble the LSM patches and observe that less overlapping produces more patch-edge artifacts and the partition of unit with a Gaussian window has the best performance. Numerical experiments for synthetic and field data show that the proposed LsmGAN method can produce high-quality images with balanced amplitudes, reduced artifacts and improved resolution.
- 14 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 61 (01962892) , 1-13
LsmGANs: Image-Domain Least-Squares Migration Using a New Framework of Generative Adversarial Networks
Compared with traditional adjoint migration, the least-squares migration (LSM) can effectively mitigate the unbalanced illumination and limited resolution associated with finite acquisition apertures, complex overburden structures and band-limited records. Data-domain LSM needs many times of Born modeling and adjoint migration to converge to a good solution, which is still challenging for large-scale 3D model under current computational capacity. To reduce computational cost and produce high-quality images, we directly approximate the Hessian inverse in the image-domain LSM using a new framework of generative adversarial networks (GANs). The migrated images, source illumination and migration velocity model are used as input data for the GANs, and the ground-truth reflectivity is utilized as the label data to train the network. Directly applying conventional GAN framework to implement the image-domain LSM leads to dislocated reflection events and incorrect images. To overcome this issue, we develop a new GAN framework that is more suitable for the Hessian approximation of image-domain LSM, which is named as LsmGANs. In the new framework, we use a max-pooling instead of convolution to downsample the feature maps to capture horizontal and vertical variations of reflectors. This enables us to map reflection events to correct location in downsampling. To address the lateral discontinuity of events in the predicted image from conventional GANs, we further apply multiple transform layers to strengthen feature transformation to guide Hessian approximation. Finally, we add the skip connection in the transform layer to enhance the information exchange of the feature channels and avoid the gradient vanishing problem to improve image resolution. Assembling predicted patches to construct a whole reflectivity image is a key step in the neural-network-based LSM. We investigate four strategies using different overlapping ratio and window functions to assemble the LSM patches and observe that less overlapping produces more patch-edge artifacts and the partition of unit with a Gaussian window has the best performance. Numerical experiments for synthetic and field data show that the proposed LsmGAN method can produce high-quality images with balanced amplitudes, reduced artifacts and improved resolution.
- 24 October 2022
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 88 (1) , WA281-WA291
Reflection and diffraction separation in the dip-angle common-image gathers using convolutional neural network
In exploration seismology, reflections have been extensively used for imaging and inversion to detect hydrocarbon and mine resources, which are generated from subsurface continuous impedance interfaces. When the interface is not continuous and its size reduces to less than half-wavelength, reflected wave becomes diffraction. Reflections and diffractions can be used to image subsurface targets, and the latter is helpful to resolve small-scale discontinuities, such as fault plane, pinch out, Karst caves, and salt edge. However, the amplitudes of diffractions are usually much weaker than that of reflections. This makes it difficult to directly identify and extract diffractions from unmigrated common-shot or common-middle-point gathers. Migrating seismic data into a subsurface location for different reflector dip angles yields a dip-angle-domain common-image gather (DACIG). One DACIG represents the migrated traces at a fixed lateral position for different reflector dips. The reflection and diffraction have different geometric characteristics in DACIG, which provides one opportunity to separate diffractions and reflections. In this study, we present an efficient and accurate diffraction separation and imaging method using a convolutional neural network (CNN). The training data set of DACIGs is generated using one pass of seismic modeling and migration for velocity models with and without artificial scatterers, respectively. Then, a simplified end-to-end CNN is trained to identify and extract reflections from the migrated DACIGs that contain reflections and diffractions. Next, two adaptive subtraction strategies are presented to compute the diffraction DACIGs and stacked images, respectively. Numerical experiments for synthetic and field data demonstrate that the proposed method can produce accurate reflection and diffraction separation results in DACIGs, and the stacked image has a good resolution for subsurface small-scale discontinuities.
- 24 October 2022
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 88 (1) , WA281-WA291
Reflection and diffraction separation in the dip-angle common-image gathers using convolutional neural network
In exploration seismology, reflections have been extensively used for imaging and inversion to detect hydrocarbon and mine resources, which are generated from subsurface continuous impedance interfaces. When the interface is not continuous and its size reduces to less than half-wavelength, reflected wave becomes diffraction. Reflections and diffractions can be used to image subsurface targets, and the latter is helpful to resolve small-scale discontinuities, such as fault plane, pinch out, Karst caves, and salt edge. However, the amplitudes of diffractions are usually much weaker than that of reflections. This makes it difficult to directly identify and extract diffractions from unmigrated common-shot or common-middle-point gathers. Migrating seismic data into a subsurface location for different reflector dip angles yields a dip-angle-domain common-image gather (DACIG). One DACIG represents the migrated traces at a fixed lateral position for different reflector dips. The reflection and diffraction have different geometric characteristics in DACIG, which provides one opportunity to separate diffractions and reflections. In this study, we present an efficient and accurate diffraction separation and imaging method using a convolutional neural network (CNN). The training data set of DACIGs is generated using one pass of seismic modeling and migration for velocity models with and without artificial scatterers, respectively. Then, a simplified end-to-end CNN is trained to identify and extract reflections from the migrated DACIGs that contain reflections and diffractions. Next, two adaptive subtraction strategies are presented to compute the diffraction DACIGs and stacked images, respectively. Numerical experiments for synthetic and field data demonstrate that the proposed method can produce accurate reflection and diffraction separation results in DACIGs, and the stacked image has a good resolution for subsurface small-scale discontinuities.