Search

The potential of blended seismic acquisition to improve acquisition efficiency and cut acquisition costs is still open, particularly with efficient deblending algorithms to provide accurate deblended data for subsequent processing procedures. In recent years, deep learning algorithms, particularly supervised algorithms, have drawn much attention over conventional deblending algorithms due to their ability to nonlinearly characterize seismic data and achieve more accurate deblended results. Supervised algorithms require large amounts of labeled data for training, yet accurate labels are rarely accessible in field cases. We present a self-supervised multistep deblending framework that does not require clean labels and can characterize the decreasing blending noise level quantitatively in a flexible multistep manner. To achieve this, we leverage the coherence similarity of the common shot gathers (CSGs) and the common receiver gathers (CRGs) after pseudo-deblending. The CSGs are used to construct the training data adaptively, where the raw CSGs are regarded as the label with the corresponding artificially pseudo-deblended data as the initial training input. We employ different networks to quantitatively characterize decreasing blending noise levels in multiple steps for accurate deblending with the help of a blending noise estimation–subtraction strategy. The training of one network can be efficiently initialized by transfer learning from the optimized parameters of the previous network. The optimized parameters trained on CSGs are used to deblend all CRGs of the raw pseudo-deblended data in a multistep manner. Tests on synthetic and field data validate the proposed self-supervised multistep deblending algorithm, which outperforms the multilevel blending noise strategy. [ABSTRACT FROM AUTHOR]

In seismic data acquisition, because of several factors, such as surface barriers, receiver failure, noise contamination and budget control, seismic records often exhibit irregular sampling in the space domain. As corrupted seismic records have a negative effect on seismic migration, inversion and interpretation, seismic trace interpolation is a key step in seismic data pre‐processing. In this paper, we propose a high‐efficiency and high‐precision seismic trace interpolation method for irregularly spatially sampled data by combining an extreme gradient boosting decision tree and principal component analysis in a semi‐supervised learning method. The adjacent trace number, sampling number and amplitudes of the effective seismic data were taken as features to build the training data set for the extreme gradient boosting decision tree. Principal component analysis is applied to remove redundant information and accelerate the training speed. This is different from the traditional trace interpolation method in that the proposed method is data‐driven; therefore, it does not require any assumptions. Compared with other deep learning‐based trace interpolation methods, the proposed method has fewer control parameters and learning labels and a smaller training cost. Experiments using synthetic and field data demonstrated the validity and flexibility of this trace interpolation method. [ABSTRACT FROM AUTHOR]

Blended acquisition is still open to improve the efficiency of seismic data acquisition. Deblending is an essential procedure to provide separated gathers for subsequent migration and inversion to characterize subsurface medium. Because of missing data, the incompleteness of blended data poses challenges for accurate deblending. Besides, the computational cost increases sharply as the increasing volume of seismic data and traditional deblending algorithms only characterize seismic data in a linear manner. To improve the deblending efficiency and accuracy, deep learning algorithms have drawn much attention to express seismic data in a nonlinear manner via supervised learning as the fast development of Graphic Processing Unit. For field cases with blending noise and sparseness contamination, we design an innovative workflow for deblending and recovery of incomplete blended data via multi-level blending noise assisted multiresolution ResUnet (MultiResUnet). The MultiResUnet combines the advantages of ResNet and Unet to characterize seismic data accurately. The designed network is trained in a supervised manner in the common receiver domain (CRD) for joint deblending and missing shot information reconstruction with multi-level blending noise to mimic the blending noise level during iterative deblending. For the reconstruction of missing receiver information, the training data in the CRD is used again to fine-tune the trained network for the pure reconstruction issue. The fine-tuned MultiResUnet is implemented for the common shot gathers (CSGs) with the projection onto convex sets strategy to reconstruct the information of missing receivers. The reconstruction accuracy can be guaranteed because the training common receiver gathers have similar features with the test CSGs. To process seismic data with different features, the trained network can be regarded as initialization based on transfer learning. Synthetic data and field data numerical examples demonstrate the validity of the proposed innovative workflow to separate blended data and reconstruct the information of missing shots and receivers. Article Highlights: Joint deblending and reconstruction of missing shots are iteratively achieved in the common receiver gather based on the designed MultiResUnet with multi-level blending noise The missing receiver information is recovered in the common shot gather based on transfer learning Numerical examples of synthetic data and field data demonstrate the validity of the proposed method in deblending and missing information recovery based on the designed deep learning workflow [ABSTRACT FROM AUTHOR]

The blended acquisition allows multiple sources to be simulated simultaneously in a narrow time interval, which can improve the acquisition efficiency and reduce the acquisition cost tremendously. However, the overlapped information from multiple sources poses challenges for traditional seismic data migration or inversion algorithms. Thus, accurate and efficient deblending should be implemented as a pre-requisite. Traditional inversion-based deblending algorithms can provide deblended data with a high computational burden, especially for a large volume of seismic data. As a deep learning strategy can match seismic data accurately in a nonlinear way through supervised learning, we propose a U-net-based accurate deblending algorithm, which incorporates transfer learning and an iterative strategy. A set of labeled synthetic data with a blending fold of 2 are classified into the training and validation data for U-net training and validation. Field data are regarded as the test data to assess the performance of the trained U-net. To guarantee the deblending performance of the field data to some extent, parts of field data with labels are used to fine-tune the trained U-net based on transfer learning. The fine-tuning procedure is relatively fast within several minutes. To further improve the deblending performance, we incorporate an iterative strategy with the fine-tuned U-net. The deblending performance is promising in the quality and computational efficiency compared with the curvelet-thresholding-based deblending method, which demonstrates the validity of the proposed intelligent deblending method. [ABSTRACT FROM AUTHOR]

Supervised classifiers play important roles in seismic interpretation. Especially in gas chimney detection, the multilayer perceptron (MLP), a supervised classifier, is widely used. However, the lack of labeled data usually limits applications of the supervised classifiers. To get more accurate seismic interpretation results, we take advantage of the information of unlabeled data by unsupervised feature extraction and semi-supervised classification methods. The sparse autoencoder (SAE) is an unsupervised learning method that can extract the features of data without labels, and the transductive support vector machine (TSVM) is a semi-supervised method that trains a classifier according to both labeled and unlabeled data. In this paper, we propose a semi-supervised learning framework that combines SAE and TSVM to detect gas chimneys. In this framework, SAE is used to extract features from data and TSVM is used to classify the labeled and unlabeled features. Therefore, the unlabeled data is taken advantage of in both unsupervised feature extraction and semi-supervised classification to improve accuracy. In order to improve the precision of detection, the attributes of neighborhood regions are also utilized. Due to the information learned from plenty of unlabeled data, the proposed framework performs well. Numerical experiments are carried out on sample sets and field data. The proposed framework has higher testing accuracy than the traditional MLP method, especially when the labeled training set is small. In field data experiments, the proposed framework also gets good prediction results for gas chimney locations. [ABSTRACT FROM AUTHOR]