Attenuation of migration artifacts on Kirchhoff migrated seismic data can be challenging due to the relatively low amplitude of migration artifacts compared to reflections as well as the overlap in the kinematics of r...Attenuation of migration artifacts on Kirchhoff migrated seismic data can be challenging due to the relatively low amplitude of migration artifacts compared to reflections as well as the overlap in the kinematics of reflection and migration smiles.Several‘conventional’filtering methods exist and recently deep learning based workflows have been proposed.A deep learning workflow can be a simple and fast alternative to existing methods.In case of supervised training of a deep neural network using training data made by physics-based modelling or actual migrations is expensive and lacks diversity in terms of noise,amplitude,frequency content and wavelet.This can result in poor generalization beyond the training data without re-training and transfer learning.In this paper we demonstrate successful applications of migration smile separation using a conventional U-net architecture.The novelty in our approach is that we do not use synthetic data created from physics-based modelling,but instead use only synthetic data build form basic geometric shapes.Our domain of application is the migrated common offset domain,or simply the stack of the pre-stack migrated data,where reflections resemble local geology and migration smiles are upward convex hyperbolic patterns.Both patterns were randomly perturbed in many ways while maintaining their intrinsic features.This approach is inspired by the common practice of data augmentation in deep learning for machine vision applications.Since many of the standard data augmentation techniques lack a geophysical motivation,we have instead perturbed our synthetic training data in ways to make more sense for a signal processing perspective or given our‘domain knowledge’of the problem at hand.We did not have to retrain the network to produce good results on the field dataset.The large variety and diversity in examples enabled the trained neural network to show encouraging results on synthetic and field datasets that were not used in training.展开更多
Estimation of good velocity models under complex near-surface conditions remains a topic of ongoing research.We propose to predict near-surface velocity profiles from surface-waves transformed to phase velocity-freque...Estimation of good velocity models under complex near-surface conditions remains a topic of ongoing research.We propose to predict near-surface velocity profiles from surface-waves transformed to phase velocity-frequency panels in a data-driven manner using deep neural networks.This is a different approach from many recent works that attempt to estimate velocity from directly reflected body waves or guided waves.A secondary objective is to analyze the influence on the prediction accuracy of various commonly employed deep learning practices,such as transfer learning and data augmentations.Through numerical experiments on synthetic data as well as a real geophysical example,we demonstrate that transfer learning as well as data augmentations are helpful when using deep learning for velocity estimation.A third and final objective is to study lack of generalization of deep learning models for out-of-distribution(OOD)data in the context of our problem,and present a novel approach to tackle it.We propose a domain adaptation network for training deep learning models that uses a priori knowledge on the range of velocity values in order to constrain mapping of the output.The final comparison on field data,which was not part of the training data,show the deep neural network predictions compare favorably with a conventional velocity model estimation obtained with a dispersion curve inversion workflow.展开更多
Geophysicists interpreting seismic reflection data aim for the highest resolution possible as this facilitates the interpretation and discrimination of subtle geological features.Various deterministic methods based on...Geophysicists interpreting seismic reflection data aim for the highest resolution possible as this facilitates the interpretation and discrimination of subtle geological features.Various deterministic methods based on Wiener filtering exist to increase the temporal frequency bandwidth and compress the seismic wavelet in a process called spectral shaping.Auto-encoder neural networks with convolutional layers have been applied to this problem,with encouraging results,but the problem of generalization to unseen data remains.Most published works have used supervised learning with training data constructed from field seismic data or synthetic seismic data generated based on measured well logs or based on seismic wavefield modelling.This leads to satisfactory results on datasets similar to the training data but requires re-training of the networks for unseen data with different characteristics.In this work seek to improve the generalization,not by experimenting with network architecture(we use a conventional U-net with some small modifications),but by adopting a different approach to creating the training data for the supervised learning process.Although the network is important,at this stage of development we see more improvement in prediction results by altering the design of the training data than by architectural changes.The approach we take is to create synthetic training data consisting of simple geometric shapes convolved with a seismic wavelet.We created a very diverse training dataset consisting of 9000 seismic images with between 5 and 300 seismic events resembling seismic reflections that have geophysically motived perturbations in terms of shape and character.The 2D U-net we have trained can boost robustly and recursively the dominant frequency by 50%.We demonstrate this on unseen field data with different bandwidths and signal-to-noise ratios.Additionally,this 2D U-net can handle non-stationary wavelets and overlapping events of different bandwidth without creating excessive ringing.It is also robust in the presence of noise.The significance of this result is that it simplifies the effort of bandwidth extension and demonstrates the usefulness of auto-encoder neural network for geophysical data processing.展开更多
文摘Attenuation of migration artifacts on Kirchhoff migrated seismic data can be challenging due to the relatively low amplitude of migration artifacts compared to reflections as well as the overlap in the kinematics of reflection and migration smiles.Several‘conventional’filtering methods exist and recently deep learning based workflows have been proposed.A deep learning workflow can be a simple and fast alternative to existing methods.In case of supervised training of a deep neural network using training data made by physics-based modelling or actual migrations is expensive and lacks diversity in terms of noise,amplitude,frequency content and wavelet.This can result in poor generalization beyond the training data without re-training and transfer learning.In this paper we demonstrate successful applications of migration smile separation using a conventional U-net architecture.The novelty in our approach is that we do not use synthetic data created from physics-based modelling,but instead use only synthetic data build form basic geometric shapes.Our domain of application is the migrated common offset domain,or simply the stack of the pre-stack migrated data,where reflections resemble local geology and migration smiles are upward convex hyperbolic patterns.Both patterns were randomly perturbed in many ways while maintaining their intrinsic features.This approach is inspired by the common practice of data augmentation in deep learning for machine vision applications.Since many of the standard data augmentation techniques lack a geophysical motivation,we have instead perturbed our synthetic training data in ways to make more sense for a signal processing perspective or given our‘domain knowledge’of the problem at hand.We did not have to retrain the network to produce good results on the field dataset.The large variety and diversity in examples enabled the trained neural network to show encouraging results on synthetic and field datasets that were not used in training.
文摘Estimation of good velocity models under complex near-surface conditions remains a topic of ongoing research.We propose to predict near-surface velocity profiles from surface-waves transformed to phase velocity-frequency panels in a data-driven manner using deep neural networks.This is a different approach from many recent works that attempt to estimate velocity from directly reflected body waves or guided waves.A secondary objective is to analyze the influence on the prediction accuracy of various commonly employed deep learning practices,such as transfer learning and data augmentations.Through numerical experiments on synthetic data as well as a real geophysical example,we demonstrate that transfer learning as well as data augmentations are helpful when using deep learning for velocity estimation.A third and final objective is to study lack of generalization of deep learning models for out-of-distribution(OOD)data in the context of our problem,and present a novel approach to tackle it.We propose a domain adaptation network for training deep learning models that uses a priori knowledge on the range of velocity values in order to constrain mapping of the output.The final comparison on field data,which was not part of the training data,show the deep neural network predictions compare favorably with a conventional velocity model estimation obtained with a dispersion curve inversion workflow.
文摘Geophysicists interpreting seismic reflection data aim for the highest resolution possible as this facilitates the interpretation and discrimination of subtle geological features.Various deterministic methods based on Wiener filtering exist to increase the temporal frequency bandwidth and compress the seismic wavelet in a process called spectral shaping.Auto-encoder neural networks with convolutional layers have been applied to this problem,with encouraging results,but the problem of generalization to unseen data remains.Most published works have used supervised learning with training data constructed from field seismic data or synthetic seismic data generated based on measured well logs or based on seismic wavefield modelling.This leads to satisfactory results on datasets similar to the training data but requires re-training of the networks for unseen data with different characteristics.In this work seek to improve the generalization,not by experimenting with network architecture(we use a conventional U-net with some small modifications),but by adopting a different approach to creating the training data for the supervised learning process.Although the network is important,at this stage of development we see more improvement in prediction results by altering the design of the training data than by architectural changes.The approach we take is to create synthetic training data consisting of simple geometric shapes convolved with a seismic wavelet.We created a very diverse training dataset consisting of 9000 seismic images with between 5 and 300 seismic events resembling seismic reflections that have geophysically motived perturbations in terms of shape and character.The 2D U-net we have trained can boost robustly and recursively the dominant frequency by 50%.We demonstrate this on unseen field data with different bandwidths and signal-to-noise ratios.Additionally,this 2D U-net can handle non-stationary wavelets and overlapping events of different bandwidth without creating excessive ringing.It is also robust in the presence of noise.The significance of this result is that it simplifies the effort of bandwidth extension and demonstrates the usefulness of auto-encoder neural network for geophysical data processing.
