We leveraged an unsupervised deep learning (DL) network for blind noise attenuation in a massive multi-input multi-output (MIMO) system without prior knowledge or labeled data. Initially, the real and imaginary parts of the received signal are concatenated in a 2-channel matrix, which is then partitioned into antenna-group patches to improve denoising efficiency. Afterward, our proposed gated recurrent unit (GRU) U-Net (GRUU-Net) network is applied to patch-wise real/imaginary data. The GRUU-Net network consists of gated recurrent units arranged in a U-Net architecture; the encoder extracts salient features from the input, while the decoder reconstructs the input patches. The main idea of GRUU-Net is that it combines the structural denoising capability of U-Net with the sequence modeling strength of GRUs and incorporates attention mechanisms to enhance feature integration within an iterative architecture. The numerical evaluation shows that our proposed model outperforms the benchmark methods across different signal-to-noise ratios (SNRs) and numbers of base station antennas (BSs). Additionally, we confirmed that applying the model simultaneously across multiple subsystems, where each patch can be considered a subsystem serving the same users in the same environment, results in better performance.

This work proposes a deep learning (DL) architecture for estimating full-field deformation maps from ground-based complex radar data. Unlike existing methods that focus on estimating deformation time series at selected points, the proposed model directly estimates the entire deformation map from a single-frame complex radar observation. This capability enables faster, spatially comprehensive monitoring, which is crucial for real-time early warning systems (EWSs) in Internet of Things (IoT)-enabled environments. The primary input to the model is a 2-channel matrix formed by the real and imaginary components of the complex radar data. We propose a dual cross-attentive recurrent U-Net architecture (DCRU-Net) consisting of two networks trained simultaneously. The two networks, a primary and an auxiliary network, are based on long short-term memory (LSTM) units arranged in a U-Net architecture, where the model inputs are spatially restructured into sequences, enabling the model to learn deformation-related phase variations across the scene rather than modeling temporal dependencies as in conventional approaches. The primary network processes the 2-channel complex radar data, which is the leading network of the model. In contrast, the auxiliary network is guided by the amplitude dispersion index (ADI) to emphasize coherent scattering regions during training. A cross-attention mechanism enables interaction between the two networks, allowing the auxiliary network to guide the primary network toward physically meaningful regions and suppress clutter. The motivation behind this architecture is to predict the main features of the deformation maps while minimizing the impact of slight variations, which can result in clutter-induced distortions. The model is evaluated against benchmark methods, and the results demonstrate that our model outperforms existing methods and has better alignment with the ideal deformation maps, showing its potential for real-world radar applications.

Abstract High-resolution remote sensing (HRRS) image scene classification has gained increasing importance in recent years, with convolutional neural networks (CNNs) showing particular promise due to their proficiency in extracting spatial features. However, traditional CNNs face significant limitations. Specifically, they struggle to capture complex semantic relationships between objects at varying scales, and they lack the ability to effectively capture long-distance dependencies between features. This limitation is especially problematic in HRRS images, where spatial relationships and semantic content are deeply intertwined. Additionally, traditional CNNs are limited in handling substantial intra-class variation and inter-class similarity, which are common in remote sensing images. To overcome these challenges, we introduce a novel Residual Channel-attention (RCA) network for scene classification. The RCA network introduces a lightweight residual structure to better capture multi-scale spatial features and incorporates a channel attention mechanism that selectively emphasizes relevant feature channels while suppressing irrelevant ones. To further refine the focus on critical image features, we integrate a squeeze-and-excitation (SE) mechanism as a self-attention component, which helps the network prioritize the most informative features and ignore background noise. We evaluated the RCA network on three public datasets: RSSCN7, PatternNet, and EuroSAT, achieving classification accuracies of 97%, 99%, and 96%, respectively. The results demonstrate that superior of the RCA network compared to state-of-the-art strategies in remote sensing image classification. Furthermore, visualization using the Grad-CAM++ algorithm highlights the effectiveness of our channel attention mechanism and underscores the RCA network’s robust feature representation capabilities.

Distributed acoustic sensing (DAS) is emerging in seismic monitoring due to its ultra-dense spatial sampling, durability to harsh environments, and sensitivity to weak ground vibration. Compared with traditional nodal geophones that are normally sparsely distributed, DAS offers unprecedented detectability for small-magnitude earthquake events, very subtle reservoir dynamics, and other weak signals among various applications. The appealing detectability of weak signals is compromised by the terabyte-scale daily continuous record that causes prohibitive storage problems. The current solution is to save only the segmented data of interest, e.g., a certain length around a target event. Here, we tackle the urgent storage problem of DAS monitoring by designing a deep-learning (DL) based compression algorithm. The compression algorithm can be split into two major components. The first part is the encoder based on the vision transformer architecture, where the input multi-channel DAS dataset goes through an encoding process to output the key features from the input. The second part is the decoder, where the features are optimally combined to reconstruct the data of the original scale. The optimal network parameters are obtained via an unsupervised training process, aiming at minimizing the difference between the reconstructed and input data. In the proposed DL-based compression algorithm, only the decoder’s weight parameters and extracted features from the input data through the encoder are saved on the disk, which is sufficient to reconstruct a high-fidelity dataset. The proposed compression algorithm can reach around 50 times the compression rate for a gigabyte-scale DAS dataset without unsatisfactory reconstruction performance.

Image enhancement is crucial for improving the resolution of seismic images obtained from band-limited data. While machine learning techniques, particularly the U-Net model, have shown significant progress in this area, they often require substantial computational resources and time. To address these challenges, we introduce a transformer-based approach for enhancing seismic image resolution, which incorporates convolutional layers, an average pooling layer, and an efficient transformer (ET). The ET leverages efficient multihead attention (EMHA) to capture long-term dependencies among image blocks, focusing on the pixels within their contextual surroundings. In our proposed model, we use a combined loss function consisting of the mean square error (mse) and the structural similarity (SSIM) to enhance the network’s learning capability. By training the model on synthetic seismic data, we observe improved structural features, enhanced resolution, and effective denoising. Notably, our approach outperforms the U-Net model in terms of SSIM and the peak signal-to-noise ratio (SNR). Furthermore, we evaluate the pretrained model on several field datasets, yielding promising results compared to the benchmark method. This demonstrates the potential applicability and effectiveness of our proposed approach in real-world scenarios.

The focal mechanism provides seismological constraints on the geological faults that generate the earthquakes and thus is important for regional seismotectonic research. Focal mechanism calculation based on the P-wave first-motion-polarity is a widely used method, particularly helpful for small to moderate-size earthquakes. However, determining the P-wave first-motion polarity can be challenging and subjective for smaller earthquakes. Here, we propose a deep-learning method (EQpolarity) for determining the P-wave first-motion polarity using the vertical-component seismic waveforms. The proposed deep-learning method was trained using a large-scale dataset from South California and then adapted to the Texas earthquake data via a transfer learning method. The original and secondary models obtained 95.43% and 98.82% accuracy on the Texas database, respectively, indicating the effectiveness of transfer learning. We further apply the deep learning method to thousands of events on the TexNet catalog to determine the focal mechanisms. Most of the focal mechanism solutions align well with the strikes, dips, and rakes of the known faults that were explored previously using full-waveform-based methods. The generation of the large focal mechanism database offers significant insights into the seismotectonic status of West Texas. The open-source package of EQpolarity can be accessed at https://github.com/chenyk1990/eqpolarity.

Predicting the peak ground acceleration from the first few seconds after the P-wave arrival time is crucial in estimating the ground motion intensity of the earthquake. The early estimation of peak ground acceleration supports the earthquake early warning system to generate the warning. Here, we propose to use the vision transformer to predict the peak ground acceleration using 4-sec three-channel single-station seismograms, i.e., 1s prior to the P-wave arrival and 3s subsequent to the arrival. The vision transformer can significantly extract remarkable information from the data resulting in superior prediction performance. The core layer of the vision transformer is the multi-head attention network which highlights the significant features of the input data. We train and evaluate the proposed algorithm using the Italian earthquake waveform data, where the proposed algorithm shows a promising result. The proposed vision transformer network utilizes an augmentation strategy to improve the learning ability of the model. Our proposed method is compared to the benchmark deep learning methods and empirical ground-motion models and outperforms all of them. The proposed algorithm can also predict the peak ground acceleration accurately using only 2-sec data after the P-wave arrival time. The proposed vision transformer architecture can also be integrated into a peak ground acceleration classification framework. Finally, the proposed algorithm is tested using real-time data and shows accurate results, indicating its applicability in real-time monitoring.

Earthquake data recorded in Texas are dramatically different from other places because of the various types of noise caused by oil and gas production or anthropogenic activities. This causes a relatively lower signal-to-noise ratio (SNR) and a strong challenge to leverage a globally trained deep learning model for earthquake detection. To combat the challenging data characteristics when monitoring seismicity using deep learning, we propose to apply transfer learning to a globally optimal phase-picking model using regional earthquake data compiled from Texas. Specifically, we first train an advanced deep learning model based on the compact convolutional transformer (EQCCT) using a global earthquake dataset. Then, we construct individual datasets from each of the main basins in Texas and apply transfer learning to each basin-scale database, intending to obtain optimal picking performance in each basin. As a result, the precision, recall, and$F1$-score significantly increased from the original EQCCT model to the fine-tuned model in the Delaware and Midland basins. The standard deviations of the picking errors of both P- and S-wave phases accordingly decrease significantly. The greatly improved EQCCT models help detect more P- and S-wave arrivals, facilitating a more successful association and location. Transfer learning models using Texas data and Texas basin-based transfer learning models with detailed documentation can be downloaded fromhttps://github.com/omarmohamed15/Picking-Texas/tree/main.

Salt body segmentation is a critical part of structural interpretation and oil and gas exploration for subsalt reservoirs. Existing automatic salt body segmentation techniques mostly use supervised learning strategies. It is challenging to generate a large number of labels by manual labeling, especially for 3-D salt bodies. Here, we propose a self-supervised learning (SSL) framework called Salt3DNet, for 3-D salt body segmentation. This framework is divided into two stages: pretraining and fine-tuning of downstream tasks. In the pretraining stage, we use the Barlow twins (BTs) method to pretrain the encoder and reduce redundancy in a contrastive learning manner to learn high-level data representations. In the fine-tuning stage, we construct two encoders to reconstruct 3-D seismic data and segment salt bodies in a multitask collaborative learning way. The encoder and decoder are composed of the 3-D fully convolutional DenseNet and soft attention mechanism, where the latter represents the selective kernel block (SKB) with multiple kernels of different sizes. Salt3DNet calculates the correlation matrix of features from different perspectives in the pretraining stage and makes it close to the identity matrix to obtain a more prosperous feature representation. Then, Salt3DNet uses a limited number of labeled samples for training. According to the evaluation metrics, the proposed network has demonstrated promising salt segmentation performance in 3-D SEG advanced modeling (SEAM) synthetic data and$F3$block real seismic data. In addition, the proposed network is demonstrated to have higher prediction accuracy than state-of-the-art salt segmentation frameworks through ablation experiments.

We present a novel deep-learning method using the ConvMixer network for automatic earthquake location. The proposed ConvMixer network utilizes three-component waveform recordings of single stations for estimating the hypocenter location. The ConvMixer network is a patch-based architecture that combines depthwise and pointwise convolutions to extract the global and local information of the earthquake waveforms. We train and test the proposed method using the Italian seismic dataset (INSTANCE). The ConvMixer network estimates the earthquake hypocenter locations with high accuracy, reaching a mean absolute error (MAE) of 2.71 km for the epicenter distance, and 1.15 km for the depth. In addition, we use the global STanford EArthquake Dataset (STEAD) to further evaluate the performance of the ConvMixer. As a result, the ConvMixer network achieves MAEs of 2.27 km and 1.19 km for the distance and the depth, respectively. The proposed ConvMixer network is compared to the benchmark methods, i.e., ResNet, AlexNet, MobileNet, and Xception, and outperforms all of them.

Passive seismic source location imaging is important to various scientific and engineering research topics spanning from unconventional reservoir development in exploration seismology to seismic hazard prevention in the earthquake seismology community. The emerging machine-learning (ML) techniques enable the location of passive seismic sources with unprecedented efficiency and accuracy. Most of the state-of-the-art ML methods are based on waveforms, as required by the most popular convolutional neural network (CNN) architecture, which is prone to the sensitivity of velocity models. Here, we present a traveltime-based ML method, RFloc3D, to locate passive seismic sources from manually or automatically picked P- and S-wave arrivals. The proposed method is similar to traditional traveltime-based location methods, where the inverse mapping from arrival times to the passive source location is obtained by inverting a nonlinear inverse problem, but differs in leveraging the random forest (RF) method to learn the inverse mapping relation from numerous eikonal-based forward simulations. Details and analyses of the proposed RFloc3D method are illustrated based on a microseismic monitoring setup. Numerical and real data examples show that the proposed method is capable of real-time location. The inclusion of S-wave arrivals, most importantly, the differential time between P- and S-wave arrivals, helps significantly to reduce the depth error (e.g., decreasing the mean absolute error (MAE) to a half) of the located sources.

We propose to implement a compact convolutional transformer (CCT) for picking the earthquake phase arrivals (EQCCT). The proposed method consists of two branches, with each of them responsible for picking the arrival times of the P- or S-wave phases. We use the STEAD dataset to train and validate the proposed EQCCT algorithm. We split the STEAD dataset into 85% for training, 5% for validation, and 10% for testing To facilitate the training process, we implement several data augmentation strategies to the training set by adding Gaussian noise, randomly shifting the waveforms, adding a second earthquake to the input window, and dropping one or two channels from the seismogram in the STEAD dataset. As a result, the EQCCT model outperforms both EQTransformer and PhaseNet, the two most popular deep-learning-based phase-picking methods. Considering the true positive criterion as the picked phases arriving within 0.5 s of the reference times, the EQCCT achieves the lowest mean absolute error (MAE) compared to the EQTransformer and PhaseNet methods for the STEAD, Japanese, Instance and Texas datasets. Our EQCCT network also demonstrates superior performance in other metrics such as precision, recall, and F1 score. We apply the pre-trained model to three independent datasets (not included in the training set), i.e., the Japanese, Texas, and Instance datasets, and achieve higher picking accuracy than the EQTransformer and the PhaseNet in terms of various statistical metrics, demonstrating a stronger robustness and generalization ability of the EQCCT. The real-time application of EQCCT in the Texas Seismological Network (TexNet) further demonstrates its production-ready performance in terms of detection and phase-picking accuracy.

3-D salt segmentation is important for many research topics spanning from exploration geophysics to structural geology. In seismic exploration, 3-D salt segmentation is directly related to the velocity modeling building that affects many processing steps, such as seismic migration and full waveform inversion. Manually picking the salt boundary becomes prohibitively time-consuming when the data size is too large. Here, we develop a highly generalized fully convolutional DenseNet for automatic salt segmentation. A squeeze-and-excitation network is used as a self-attention mechanism for guiding the proposed network to extract the most significant information related to the salt signals and discard the others. The proposed framework is a supervised technique and shows robust performance when applied to a new dataset using transfer learning and a small amount of training data. We test the robustness of the proposed framework on the Kaggle TGS salt segmentation dataset. To demonstrate the generalization ability of the framework, we further apply the trained model to an independent dataset synthesized from the 3-D SEAM model. We apply transfer learning to finely tune the trained model from the TGS dataset using only a small percentage of data from the 3-D SEAM dataset and obtain satisfactory results.

Accurate estimation of reservoir parameters (e.g., permeability and porosity) helps to understand the movement of underground fluids. However, reservoir parameters are usually expensive and time-consuming to obtain through petrophysical experiments of core samples, which makes a fast and reliable prediction method highly demanded. In this article, we propose a deep learning model that combines the 1-D convo- lutional layer and the bidirectional long short-term memory network to predict reservoir permeability and porosity. The mapping relationship between logging data and reservoir parameters is established by training a network with a combination of nonlinear and linear modules. Optimization algorithms, such as layer normalization, recurrent dropout, and early stopping, can help obtain a more accurate training model. Besides, the self-attention mechanism enables the network to better allocate weights to improve the prediction accuracy. The testing results of the well-trained network in blind wells of three different regions show that our proposed method is accurate and robust in the reservoir parameters prediction task.

We develop a random forest (RF) model for rapid earthquake location with an aim to assist earthquake early warning (EEW) systems in fast decision making. This system exploits P-wave arrival times at the first five stations recording an earthquake and computes their respective arrival time differences relative to a reference station (i.e., the first recording station). These differential P-wave arrival times and station locations are classified in the RF model to estimate the epicentral location. We train and test the proposed algorithm with an earthquake catalog from Japan. The RF model predicts the earthquake locations with high accuracy, achieving a mean absolute error (MAE) of 2.88 km. As importantly, the proposed RF model can learn from a limited amount of data (i.e., 10% of the dataset) and much fewer (i.e., three) recording stations and still achieve satisfactory results (MAE < 5 km). The algorithm is accurate, generalizable, and rapidly responding, thereby offering a powerful new tool for fast and reliable source-location prediction in EEW.

Discrimination between earthquakes and quarry blasts is crucial for precise seismic analysis, e.g., seismic hazard mitigation, earthquake cataloging, etc. However, the discrimination process is challenging due to the similarity of waveforms between the local earthquakes and quarry blasts. We propose to use the scalogram and the capsule neural network to distinguish between earthquakes and quarry blasts. First, we obtain the scalogram for 60s 3-channel waveforms, where we extract 10s before and 50s after the first arrival time of the seismic event. Secondly, we utilize the capsule neural network to extract the important information from the input scalogram which leads to robust classification performance. The proposed capsule neural network consists of the convolutional layer, primary capsule layer, and digit caps layer. The convolutional layer extracts the important information from the input data, and the primary capsule layer extracts the spatial relationship between different feature maps. Thirdly, we use the dynamic routing process to connect the primary capsule to the digit caps layer. We train and test the proposed capsule network using a small and unbalanced dataset which is recorded by the Egyptian Seismic Network (ENSN) in the Red Sea and the surrounded area in Egypt. Accordingly, the proposed method achieves a test accuracy of 96.08%. The proposed method is compared to the benchmark methods, i.e., convolutional neural network (CNN), AlexNet, VGG, and ResNet networks, and demonstrated to outperform all of the competing methods. Finally, we apply the proposed method to classify real-time seismic events and obtain promising results.

Diffractions in the seismic data are associated with the small-scale subsurface structures, thus their separation and imaging are helpful in characterizing the underground discontinuities with a high resolution that cannot be reached by traditional reflection imaging methods. Traditional seismic slope-based diffraction separation methods are strongly affected by the accuracy and stability of the slope estimation methods, e.g., the plane-wave destruction (PWD) method. When the local seismic slope is not properly estimated, the separated seismic diffraction waves suffer from the mixture between the reflection and diffraction energy due to their coupling in the slope map. We propose an automatic local rank-reduction (LRR) method to separate 3D diffraction waves from zero-offset seismic data, based on which we conduct 3D migration to output the diffraction images. Due to the difficulty in choosing the rank in each local 3D window, we apply an adaptive strategy to obtain the optimal rank. The proposed LRR method with adaptively selected ranks (LRRA) is applied to several 3D synthetic and field data examples and demonstrated to perform better than the traditional PWD, the LRR, and the global rank-reduction (GRR) methods.

The footprint noise is a type of coherent noise that arises from the acquisition geometry. The footprint noise is commonly seen in 3-D seismic volumes and greatly affects the amplitude-based processing and interpretation steps in the seismic exploration workflow. Thus, removal of the footprint noise is necessary for warranting a reliable interpretation output of seismic data processing. However, because footprint noise is usually weak and also spatially coherent, it is inevitable to cause damages to useful signals during its removal steps. Here, we propose a dictionary learning (DL)-based method to effectively remove the footprint noise. We design an algorithm framework to effectively learn the dictionary atoms of the signal waveforms and separate the features of the footprint noise from the learned atoms. Considering the special features of the footprint noise in the dictionary atoms, we propose a statistics-guided way to separate the dictionary atoms into footprint-affected and footprint-free atoms. Then, the footprint-affected atoms are processed via a 2-D median filtering step. The combination between the untouched footprint-free atoms and filtered footprint-affected atoms result in a better dictionary of the signal waveforms and the footprint atoms. We use residual DL to encode the input data by a linear combination of signal atoms and footprint atoms. Removal of footprint atoms and their corresponding sparse coefficients leads to a successful footprint removal. We use both 3-D synthetic and field data examples to demonstrate the effectiveness of the proposed method.

We propose a novel deep learning method using an attention-based fully convolutional dense network (FCDNet) for automatic earthquake detection. The FCDNet consists of encoder-decoder parts with skip connections, where each encode-decoder block contains a block of densely connected layers to enhance the feature learning capability. The spatial attention mechanism is added within the FCDNet to assign greater attention to useful features and hence improve the accuracy of earthquake detection. The time-frequency representations of three-component seismograms produced by the Stockwell transform are used for better extracting the hidden data features. The attention-based FCDNet extracts the time-frequency features needed for distinguishing the seismic signal from the background noise. We evaluate the performance of the proposed method using a Mediterranean dataset. The attention-based FCDNet is trained using 90% of the Mediterranean dataset and tested using the remaining 10%. Accordingly, the training and testing accuracies are 97.71% and 97.02%, respectively. The intersection over union (IoU), precision, recall, and F1-score of the attention-based FCDNet are 93.80%, 99.72%, 99.55%, and 99.64%, respectively. Moreover, to evaluate the generalization ability of the trained model, we utilize 100,000 seismic waveforms recorded in different seismic regions from the global STanford EArthquake Dataset (STEAD) dataset for testing, which shows robust performance. We also apply the attention-based FCDNet to the Japanese seismic data and compare the performance to the CRED and SCALODEEP methods. The attention-based FCDNet outperforms the benchmark methods and achieves a higher detection accuracy of 99.46%. The attention-based FCDNet is additionally evaluated using one-day continuous seismic data recording a seismic swarm that occurred in the Helike region. As a result, the attention-based FCDNet recognizes 135 earthquakes and raises 15 false alarms with a detection accuracy of 90.06%.

We develop a capsule neural network (CapsPhase) for seismic data classification and picking. CapsPhase consists of several layers, e.g., convolutional, primary capsule, and digit capsule layer. The convolutional layer extracts the significant features from the seismic data, while the primary capsule combines the extracted features into several vector representations named capsules. Afterward, the primary capsule is connected to the digit capsule layer using a dynamic routing strategy to obtain the vector representation of each output class, i.e.,$P$-wave,$S$-wave, and noise class. CapsPhase is trained using 90% of the Southern California seismic dataset, which contains 4.5 million 4 s-three-component seismograms, and is validated and tested using the remaining 10%. Accordingly, the training accuracy reaches 98.70%, while the validation accuracy is 98.67% and the testing accuracy is 98.66%. Furthermore, the CapsPhase is tested using 300 000 earthquake waveforms recorded worldwide from the STanford EArthquake Dataset (STEAD). Accordingly, the precision, recall, and F1-score of the$P$-picks corresponding to the CapsPhase reach 94.50%, 99.86%, and 97.10%, respectively, whereas the precision, recall, and F1-score of the$S$-picks corresponding to the CapsPhase are 88.05%, 99.87%, and 93.60%, respectively. In addition, CapsPhase is evaluated using the Japanese seismic data and is compared to benchmark methods, e.g., short-time average/long-time average (STA/LTA), generalized phase detection (GPD), and CapsNet methods. As a result, CapsPhase reaches F1-scores of 99.10% and 98.64% for$P$-wave and$S$-wave arrival times, respectively, and outperforms the benchmark methods. The results show that the CapsPhase has the ability to pick the arrival times accurately despite the existence of strong background noise, e.g., the signal-to-noise-ratio (SNR) can be as low as −4.97 dB. Besides, the CapsPhase detects the arrival time when the earthquake has a small local magnitude, e.g., as low as$0.1~M_{L}$. In addition, we find that the proposed algorithm has the ability to train using a small dataset, which is valuable for regions that have limited training data.

We propose to use unsupervised deep learning (DL) and attention networks to mute the unwanted components of the single-channel earthquake data. The proposed algorithm is an unsupervised technique that does not require any prior information about the input data, i.e., no need for the labeled data. The imaginary and real parts of the short-time frequency transform (STFT) are divided into several overlapped patches to be the input of the proposed DL network, while the output target is the absolute value of the STFT. The proposed DL network utilizes a customized loss function to reconstruct the signal mask, where the STFT components related to the seismic noise are muted. An adaptive thresholding technique is utilized to obtain the binary mask, which is multiplied by the real and imaginary parts of the input seismic data. The binary mask has zero values for the samples corresponding to the unwanted components and ones for the seismic signal components. Then, inverse STFT is used to reconstruct the denoised signal. The proposed algorithm is evaluated using samples from the STanford EArthquake Dataset (STEAD) and the results are compared to the benchmark denoising method, i.e., DeepDenoiser. As a result, the proposed algorithm shows a robust denoising performance and outperforms the DeepDenoiser method by 1.95 dB in terms of signal-to-noise ratio.

We develop a deep learning framework based on deep image prior (DIP) and attention networks for 3-D seismic data enhancement. First, the 3-D noisy data are divided into several overlapped patches. Second, the DIP network has a U-NET architecture, where the input patches are encoded to extract the significant latent features, while the decoder tries to reconstruct the input patches using these extracted features. Besides, the attention network is used to scale the extracted features from the encoder and the decoder. Third, the attention network output of the encoder is concatenated with that of the decoder to obtain high-order features and guide the network to extract the most significant information related to the seismic signals and discard the others. Finally, the 3-D seismic data are reconstructed using the output patches obtained by the DIP network. The proposed algorithm is an iterative and unsupervised approach, which does not require labeled data. We evaluate the proposed algorithm using several synthetic and field data examples. As a result, the proposed algorithm shows the ability to enhance the 3-D seismic data by attenuating the random noise and preserving the 3-D seismic signal with minimal signal leakage. Moreover, the proposed algorithm shows good denoising performance when tested using various types of events, e.g., linear, hyperbolic, low and high dominant frequencies, and weak amplitude. In addition, the proposed method outperforms the predictive filtering (PF) and damped rank-reduction (DRR) methods. To further understand the principle of the proposed method inside the DIP network, we analyze the weighting matrices in the encoder and decoder parts in detail. We attribute the denoising ability of the DIP network to the improvement of the extracted basis features from the encoder to the decoder layers through a deep network.

The unsplit complex frequency-shifted perfectly matched layer (CFS-PML) has been widely used in the first-order wave equation in velocity and stress while rarely formulated for the wave equation recast as a second-order system in displacement. Among different variants of PML, the unsplit CFS-PML for the second-order wave equation enjoys better absorbing performance and numerical stability, compared to the traditional PML, due to the presence of the general form of CFS stretching factors, as well as higher computational efficiency over the split PMLs since it avoids wave equation order reduction and splitting the state variables into multiple directional components. This study aims to develop an unsplit CFS-PML scheme for the second-order wave equation and devote specific attention to fractional viscoacoustic simulation where fractional time derivatives are involved and hard to be reformulated into a first-order form. In the complex space, PML is typically regarded as an analytical continuation of the real coordinates; thus, we define an explicit coordinate stretching operator acting on the Laplacian operator. This stretching operator consists of several convolution terms; each of them can be efficiently resolved by a recursive convolution updating strategy. Viscoacoustic simulations on homogeneous Pierre Shale, Marmousi model, and 3-D SEG/EAGE overthrust model verify the feasibility and absorbing the performance of our proposed scheme.

Effective random noise attenuation is critical for subsequent processing of seismic data, such as velocity analysis, migration, and inversion. Thus, the removal of seismic random noise with an uncertainty level is meaningful. Attenuating 3-D random noise in a supervised way based on deep learning (DL) is challenging because clean labels are difficult to obtain. Therefore, it is necessary to develop an adaptive unsupervised-based method for random noise attenuation. In this article, we propose a deep-denoising unsupervised learning (DDUL) network to attenuate random noise in 2-D/3-D seismic data. A patching technique is used to split 2-D/3-D seismic data into several patches to be fed into the network, which helps to expand the number of samples for training. We use the fully symmetrical structure of the autoencoder to construct the network. In each corresponding encoder and decoder layer, skip connections are added to enhance the learning of seismic data features. We construct three blocks to extract waveform features in seismic data, i.e., encoder, decoder, and skip blocks. Among them, the skip is connected between the encoder and decoder blocks of each hidden layer. The use of multiple blocks not only improves the network’s ability to extract seismic data features but also solves the problem of excessive training parameters caused by hidden layer stacking. Five 2-D/3-D synthetic and field seismic datasets are used to test the denoising performance of our proposed method. The denoising results demonstrate that our proposed method has good signal-preserving and noise attenuation capabilities in real-world applications.

Magnitude determination of earthquakes is a mandatory step before an earthquake early warning (EEW) system sends an alarm. Beneficiary users of EEW systems dependon how far they are located from such strong events. Therefore,determining the locations of these shakes is an important is sue for the tranquility of citizens as well. In light of that, this article proposes a magnitude, location, depth, and origin timecategorization using earthquake Ml magnitudes between 2 and 9.The dataset used is the fore and aftershocks of the great Tohokuearthquake of March 11,2011, recorded by three stations fromthe Japanese Hi-net seismic network. The proposed algorithmdepends on a convolutional neural network (CNN) which hasthe ability to extract significant features from waveforms thatenabled the classifier to reach a robust performance in the required earthquake parameters. The classification accuracies ofthe suggested approach for magnitude, origin time, depth, andlocation are 93.67%,89.55%,92.54%,and 89.50%, respectively.

Earthquake detection is an essential step in observational earthquake seismology. We propose to utilize a capsule neural network (CapsNet) to automatically identify and detect earthquakes. CapsNet is the new generation of deep learning architecture. It has the capability of learning with a great generalization performance from a small dataset. We train the CapsNet using 50% of the Southern California seismic data (2.25 million 4-s-three-component seismic windows) and use 222 395 waveforms from different seismic areas to evaluate the CpasNet performance, e.g., western United States, Europe, and Japan. As a result, the CapsNet misses 367 events and detects 217 305 events with an accuracy of 97.71%. Among these picked events, 210 498 events have an arrival time error below 0.2 s (96.86%) and 197968 waveforms with an arrival time error below 0.1 s (91.11%). The CapsNet precision, recall, and F1-score are 97.78%, 99.83%, and 98.79%, respectively. In addition, the CapsNet is tested using 100 000 60-s-three-component seismic noise waveforms. CapsNet shows a low false alarms rate of 1384, which gives the CapsNet an accuracy of 98.61%. In addition, CapsNet is tested using continuous seismic data associated with the 24-hours microearthquakes swarm that occurred in the Arkansas area. Accordingly, the CapsNet detects 221 earthquakes and releases 37 false alarms with a detection accuracy of 85.65%. CapsNet detects many microearthquakes with a small magnitude, as low as -1.3 Ml, and detects earthquakes that have a low signal-to-noise ratio (SNR), e.g., as low as -8.07 dB. The results of the CapsNet are compared to the benchmark methods, e.g., short-time average/long-time average (STA/LTA) and GPD methods. The CapsNet shows the highest picking accuracy and outperforms the benchmark methods.

Monitoring illegal explosions is mandatory for the safety of human life, environment, and protect the important buildings such as High-dam in Egypt. This kind of monitoring can be accomplished by detecting and identifying the explosions. If an illegal explosion happens such as quarry blast, an alarm should be reported to the government to take immediate action. However, the main problem is that many measured signals from received explosions are similar to earthquakes in their shape and both cannot differentiate from each other. Also, incorrect classification possibly will distort the real seismicity nature of the region. This problem motivates us to search for unique discriminating features to distinguish between earthquakes and explosions with precise accuracy. Therefore, in this paper, we propose to extract the discriminative features based on Autoencoder from the first few seconds after the P-wave arrival time of the event. The discriminative features are found to be in the first 60 samples after the arrival time of P-wave. Thus the first stage of the proposed algorithm is extracting the discriminative features via the Autoencoder. Then, softmax classifies the event based on these extracted features. The proposed algorithm achieves a classification accuracy of 98.55% when applied to 900 earthquakes and quarry blasts waveforms recorded by Egyptian National Seismic Network (ENSN).

The accurate detection of P-wave arrival time is imperative for determining the hypocenter location of an earthquake. However, precise detection of onset time becomes more difficult when the signal-to-noise ratio (SNR) of the seismic data is low, such as during microearthquakes. In this letter, a stacked denoising autoencoder (SDAE) is proposed to smooth the background noise. The SDAE acts as a denoising filter for the seismic data. In the proposed algorithm, the SDAE is utilized to reduce background noise such that the onset time becomes more clear and sharp. Afterward, a hard decision with one threshold is used to detect the onset time of the event. The proposed algorithm is evaluated on both synthetic and field seismic data. As a result, the proposed algorithm outperforms the short-time average/long-time average and the Akaike information criterion algorithms. The proposed algorithm accurately picks the onset time of 94.1% for 407 field seismic waveforms with a standard deviation error of 0.10 s. In addition, the results indicate that the proposed algorithm can pick arrival times accurately for weak SNR seismic data with SNR higher than -14 dB.