Seismic Impedance Inversion Using Fully Convolutional Residual Network and Transfer Learning

Wu, Bangyu; Meng, Delin; Wang, Lingling; Liu, Naihao; Wang, Ying

doi:10.1109/lgrs.2019.2963106

Cited by 122 publications

(34 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…• Initial earth model estimate is not required for DL [28], [36], [47], [34], [37], [38], as compared to FWI, which is sensitive to the choice of the initial model. • Transfer-Learning can be utilized to leverage learned knowledge from one geographical area to another [50], and between domains [46]. • The modularity of DL enables to process multi-modal input data such as well-logs and seismic data or gravity and electromagnetic, which facilitate joint inversion [66].…”

Section: Discussionmentioning

confidence: 99%

“…Therefore, mixing the two types of regularizers should be investigated as well as evaluating new types of regularizes. For example, Lunz et al • Robustness to Seismic Data Degradation: initial studies [36], [50] demonstrate good robustness of DL inversion to noise, yet, seismic data often suffers additionally from missing samples or missing traces. Therefore, future investigations of DL inversion in the presence of severe seismic data degradation models are required.…”

Section: Discussionmentioning

confidence: 99%

“…Wu et al [50] proposed a trace-to-trace fully convolutional residual network (FCRN) for impedance model building from pre-stack seismic data. Residual networks [25] employ skip connections, which are network shortcuts that jump over few layers, as depicted in figure 13(a).…”

Section: Semblance Velocity Analysismentioning

confidence: 99%

“…Residual networks have shown significantly better performance of very deep networks as compared to non-residual networks. The proposed architecture in [50] is composed of an input convolutional layer, followed by three residual blocks and a final convolutional layer that reconstructs the impedance model. The first convolutional layer of the FCRN was adapted form [40] and includes 16 1D kernels of 300 samples width.…”

Section: Semblance Velocity Analysismentioning

confidence: 99%

See 3 more Smart Citations

Deep Learning for Seismic Inverse Problems: Toward the Acceleration of Geophysical Analysis Workflows

Adler¹,

Araya‐Polo²,

Poggio³

2021

IEEE Signal Process. Mag.

View full text Add to dashboard Cite

Seismic inversion is a fundamental tool in geophysical analysis, providing a window into the Earth. In particular, it enables the reconstruction of large scale subsurface earth models for hydrocarbon exploration, mining, earthquakes analysis, shallow hazard assessment and other geophysical tasks. This article provides a comprehensive and timely overview of emerging datadriven Deep Learning (DL) solutions to seismic inverse problems, including velocity, impedance, reflectivity model building and seismic bandwidth extension. The article reviews seismic wave propagation and signal acquisition principles using large scale sensor arrays in offshore and inland exploration. In addition, the relations between the seismic forward and inverse problems are established, and prominent model-based solutions including seismic tomography, Full-Waveform Inversion (FWI) and multiscale FWI are explained. A primer to DL principles is presented, including empirical risk minimization, stochastic gradient-based learning, Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and regularization. The article then presents DL-based solutions addressing full seismic inversion as well as DL-assisted FWI solutions including Encoder-Decoder architectures, CNN-based networks and RNN-based networks with Long Short Term Memory and Gated Recurrent Units. Seismic signals features extraction methods such as semblance velocity analysis and seismic spectrograms are discussed, as well as batch vs. sequential data processing architectures. The article further presents advanced solutions based on Generative Adversarial Networks , and Physics-guided DL architectures that incorporate numerical geophysical modeling into the DL training loop, enabling unsupervised and semi-supervised learning.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Discussionmentioning

confidence: 99%

Section: Semblance Velocity Analysismentioning

confidence: 99%

Section: Semblance Velocity Analysismentioning

confidence: 99%

See 2 more Smart Citations

Deep Learning for Seismic Inverse Problems: Toward the Acceleration of Geophysical Analysis Workflows

Adler¹,

Araya‐Polo²,

Poggio³

2021

IEEE Signal Process. Mag.

View full text Add to dashboard Cite

show abstract

“…Existing deep learning methods can be grouped into noniterative methods and iterative methods [32]. The non-iterative deep learning methods usually refer to data-driven deep learning methods that use the deep neural network architectures to directly learn feed-forward mapping without any explicit physical knowledge, e.g., [33] built a data-driven convolutional neural network (CNN) to implement seismic impedance inversion. A main advantage of the non-iterative deep learning methods is that they can dramatically reduce the time complexity [34].…”

Section: Introductionmentioning

confidence: 99%

A Sequential Iterative Deep Learning Seismic Blind High-Resolution Inversion

Chen

Gao

et al. 2021

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

Seismic blind high-resolution inversion (BHRI) aims at retrieving the high-resolution data to characterize the stratigraphic structures in the case of an unknown seismic wavelet. However, the unknown wavelet and ill-posedness pose a great challenge to the high-resolution inversion. Regularization-based BHRI is an effective approach. However, it is sensitive to the sets of initial values, regularization terms, and regularization parameters, and suffers from computational burden problems. To address these issues, we propose a sequential iterative deep learning method shortened to SIDLM to implement a BHRI in a fast computational speed, which incorporates three learned components to sequentially invert initial high-resolution data, seismic wavelet, and final high-resolution data in an end-to-end fashion. Specifically, to mitigate the influence of initial values, a data-driven network U-Net is adopted to learn an initial highresolution data. Further, the architecture makes use of prior information encoded in the forward operator to build a new general Alternating Direction Method of Multipliers (ADMM)like iterative deep neural network which is instead of the traditional alternating iterative inversion. The proposed ADMMlike network utilizes the convolutional neural networks to learn the proximal operators to solve each sub-problems in alternating iterative inversion. Therefore, all parameters of BHRI, such as the regularization parameters and transform operator, can be implicitly learned from the training data sets in an end-toend fashion, not limited to the form of the penalty function. Finally, the synthetic and field data examples are conducted to demonstrate the effectiveness of the proposed SIDLM.

show abstract