Jincheng Xu

A prestack depth migration (PSDM) scheme was developed that compensates for absorption and dispersion in the absence of a conventional Q model. The proposed scheme followed the Kirchhoff approach. Instead of applying a commonly used Q model, an offset-related effective Q model was proposed, which only and fully determines, at each spatial location, the local migrated results independent of the effective Qs in other locations. The offset-related effective Q was approximately described at each spatial point using a quadratic polynomial. The related three parameters were determined during migration using a group of migrated gathers obtained by multiple homogeneous Q migrations. This was completed by solving a three-parameter optimization problem at each spatial point, which maximizes the bandwidth of the stacked spectra of partially and fully stacked local migrated sections inside the window around the point. The proposed approach overcame the challenge to estimate an interval Q model in depth, as in conventional anelastic PSDM. The resulting effective Q deabsorption PSDM was applied to synthetic 2D data and 3D field data. Higher-resolution imaging results were obtained.

Accurate source localization and source mechanism estimation in microseismic monitoring critically depend on precise subsurface velocity model reconstruction. Recent advances in data-driven approaches, such as neural network models, have demonstrated promising performance in seismic velocity inversion. However, purely data-driven methods are significantly constrained by their reliance on the completeness of the training dataset. Additionally, the performance of current neural network architectures remains suboptimal and requires further improvement. We propose a step-by-step training methodology that integrates the wave equation-based modeling with a modified U-shaped neural network (U-Net). This approach incorporates physical laws of wave propagation to constrain the updating process of the velocity model. The proposed method carries out the purely data-driven procedure first and then incorporates waveform modeling with finite-difference simulations. Besides, the new method defines dynamic weights within the loss function and introduces the Akaike information criterion (AIC) to detect and extract the P and S phases, which reduces the computational expenses as well as alleviates the artifacts in velocity models effectively. We conducted experiments with synthetic fault models and realistic layered models, and the results show that the wave equation-constrained method recovers the velocity model better than the purely data-driven algorithm. The approach leveraging physics-based and data-driven models can be easily adapted to other geophysical inversion tasks.

Migrated dip-angle gathers provide a visible domain for estimating Fresnel zones of migration. However, it remains a challenge to extract Fresnel zones from the dip-angle gathers with higher precisions and less workloads, especially in 3D cases. A commonly used approach is to yield a pair of 2D dip-angle gathers by stacking the 3D dip-angle gathers along the inline and crossline direction, then to extract the two 1D Fresnel zones from the resulting pair of 2D dip-angle gathers, and finally incorporate the two 1D Fresnel zones into 2D one. The deep learning-based data-driven method has proven efficient in automatically extracting 1D Fresnel zones from a 2D dip-angle gather. However, unlike the current data-driven methods that treat the extraction of each 1D Fresnel zone as an independent task, we propose a unified multi-task learning network based on the segmentation model to simultaneously extract a pair of 1D Fresnel zones from the pair of 2D dip-angle gathers. The network architecture consists of a shared encoder and multi-head decoder modules, thus leveraging shared knowledge and representations across tasks while extracting task-specific features through end-to-end training. The proposed architecture significantly reduces the total number of training parameters and accelerates the training process compared to independent networks for each task (i.e., single-task learning). Furthermore, the network’s generalizability is enhanced compared to single-task learning due to the fact that multiple tasks act as regularizations for each other. We demonstrate the method using a real 3D field data. Higher signal-to-noise ratio (SNR) migration results are obtained.

Microseismic events play a crucial role in mapping fault and fracture distributions in natural and induced earthquakes. Detecting and localizing microseismic events is challenging due to low signal-to-noise ratios. Waveform stacking imaging location is a practical approach for automatically detecting and localizing microseismic events, assuming that the traveltime-corrected seismic waveforms will stack and enhance coherently. However, coherent stack enhancement is susceptible to polarity reversal caused by the nonexplosive components of the source mechanism, which can lead to an unfocused source image, making it difficult to retrieve the optimal location accurately. In this study, we develop a new polarity correction method to address this issue based on the fact that the instantaneous phase difference between two seismic signals with opposite polarity is ±π. First, the Hilbert transform is applied to the original seismic record to obtain the instantaneous amplitude A(t) and phase φ(t). Then, a new signal is constructed by multiplying A(t) with cos[2φ(t)]. The constructed signals for different receivers have the same polarity; therefore, they can be used to refocus the source image. Furthermore, they preserve positive and negative amplitudes, which contributes to noise suppression. Synthetic tests show that our method can effectively achieve polarity correction and noise suppression, enabling a high-resolution source image. Application to real hydraulic fracturing data demonstrates that our method can detect and locate more microseismic events at the fracturing depths, suggesting its effectiveness and potential advantage in microseismic data processing. Because the polarity correction is performed in the data domain without relying on specific receiver layouts, the method is computationally efficient and could be applied to real-time microseismic monitoring across various sites, such as hydraulic fracturing and volcano monitoring.

Seismic migration produces the migrated images of subsurface media using seismic data, which is important for geophysical exploration. However, the adjoint-based migration methods may produce a blurry image, convolved by a Hessian matrix. To address this problem, we propose a Hessian-assisted iterative self-training learning method aimed at approximating the inverse Hessian matrix and deblurring the migrated image. First, we train a long-short term network using labelled images and use it as a teacher network to generate pseudo labels for the unlabelled images. Subsequently, we integrate the demigration and migration operators to identify the pseudo labels with high confidence levels and construct a dataset containing both the true and pseudo labels. The dataset is then used to train a student network with the injection of model noise into the network. Finally, we regard the student network as a new teacher and repeat the process in an iterative self-training learning framework. We demonstrate the effectiveness of our proposed method using two synthetic datasets and a field data. Compared with the supervised learning method, the proposed method exhibits superior generalization capabilities. This advantage stems from the incorporation of the demigration and migration operators, providing a valuable prior for the inverse Hessian matrix in training the model. In contrast to the model-driven least-squares migration methods, the proposed method yields higher-resolution images with significantly reduced computational costs. However, it may be less effective in recovering small-scale structures when confronted with an extremely limited number of labels.

Building a macro Q model for deabsorption migration using surface reflection data is challenging owing to interferences of the reflections resulting from stacked thin layers. The effective Q approach gives an alternative way to overcome this difficulty. However, manual processing is involved for effective Q estimation. This restricts the use of denser grids in building an inhomogeneous Q model. We therefore incorporate deep learning into the effective Q approach, thus yielding a deep learning‐based Q model building scheme. The resulting scheme improves the manual effective Q estimation by simultaneously accounting for the imaging resolution and induced noises using two networks. Moreover, most manual processing is reduced in spite of denser grids in building a 3D Q model. One of the networks used is a 1D convolutional neural network that determines the optimal upper cut‐off frequency for a selected Q with an input of multi‐channel amplitude spectra, and another is a residual neural network that determines the optimal Q for a series of Q values with an input of multi‐channel imaging sections inside the selected small window filtered under the corresponding upper cut‐off frequencies. As a result, a Q model that improves the imaging resolution in the absence of amplification of noises is gained. Transfer learning is used, thus reducing the training cost when applied to different geological targets. We test our scheme using 3D field data. Higher resolution images without induced noises are obtained by a deabsorption migration using the Q model built and compared to those obtained by the migration without absorption compensation.

Implementing Fresnel zones-based stacking is a pursuit in various seismic imaging methods. However, the estimation of Fresnel zones remains a challenge in 3-D migration. Migrated dip-angle gathers provide a visible domain for estimating Fresnel zones. An analytical estimation (i.e., a model-driven method) faces limitations in automatically estimating Fresnel zones in migrated dip-angle gathers in real-world situations due to complex reflections, nonuniform coverage, and noises. Human interaction remains necessary for precisely estimating Fresnel zones in migrated dip-angle gathers. Due to the high number of Fresnel zones in 3-D cases, interpolation is necessary to fill in the gaps between manually estimated zones. Aiming to reduce the workload of human interaction and mitigate interpolation errors, we propose a semantic segmentation model (i.e., a deep learning-based data-driven method) to estimate Fresnel zones in migrated dip-angle gathers automatically. We transform the estimation of Fresnel zones into a binary classification task of each pixel in dip-angle gathers. Instead of training the network using the entire dip-angle gather images, we train the network using patches to make the network focus on learning the detailed and general features within the patches. Our proposed network, named deep-supervised attention-UNet, is trained using a deep-supervised method along with a hierarchical hybrid loss function to segment the dip-angle gather on different scales. This approach yields superior segmentation results compared to the UNet model in qualitative and quantitative aspects. We test the efficiency and practicability of our method using a marine field dataset. The signal-to-noise ratio (SNR) of migration results obtained using the Fresnel zones estimated by our method is improved significantly.