The accuracy of the first-arrival time plays a key role in the overall seismic data processing flow. First-arrival picking is often formulated as a binary classification problem at the pixel level. This study proposes a physics-guided deep regression method to pick first-arrival times from seismic waveforms. The first-arrival times are encoded as a probability distribution map through the multiscale Gaussian kernel function, and their detection can be transformed into a continuous regression problem. Unlike conventional methods that treat physical features as additional input channels, we build a physics-driven attention module that directly incorporates physics knowledge as a guiding principle in the process of model learning. This module uses the sudden change in signal energy caused by seismic wave arrival as its benchmark and modifies it only when necessary. Through the decoupled design of the confidence map and the residual map, the regression model can fully leverage the strong guiding role of physics knowledge while flexibly activating data-driven alternatives when physical features fail. The physics-driven attention module is embedded within a hybrid network of Swin Transformers and convolutional neural networks to assist in capturing key information related to first arrivals. The physics-guided deep regression method and the existing methods are compared using multiple field seismic datasets. The results demonstrate that our method excels in enhancing the accuracy and stability of first-arrival picking.

Full waveform inversion (FWI) is an advanced velocity modeling method. FWI using active seismic data has high-resolution potential but often faces challenges dominated by cycle-skipping due to the insufficient low-frequency components. The ambient noise contains rich low-frequency components, but the inversion resolution is generally low. Joint FWI utilizing the complementary advantages of active and passive seismic data enables more accurate recovery of subsurface velocity structures. However, the conventional serial joint strategy is strongly dependent on the accuracy of passive source FWI, and complex interference from ambient noise introduces artifacts into inversion results, leading to instability of the serial strategy. To mitigate this problem, we develop a dynamically weighted parallel joint FWI. Firstly, an attenuated spatiotemporal window function combined with multidimensional deconvolution (MDD) seismic interferometry is designed to retrieve broadband body-wave Green’s functions from ambient noise. Subsequently, a weight function dynamically varying with iterations is established to integrate active and passive source data into a unified objective function framework, and the contributions of multi-source data are simultaneously utilized to construct a joint gradient for velocity reconstruction. Furthermore, we investigate the influence of parameter variations controlling the weight function on joint inversion performance, suggesting two parameter selection strategies, and verify their applicability and relative robustness. Numerical examples demonstrate that the proposed parallel joint FWI can effectively fuse the respective advantages of active seismic data and ambient noise, and robustly achieve high-resolution velocity modeling.

In seismic data acquisition, the simultaneous source technique has been widely used by virtue of its high acquisition efficiency. After collecting a large amount of simultaneous source data, the simultaneous source data needs to be deblended. Nevertheless，the highly coherent and intricate entanglement of aliased signals with desired signals poses a significant hurdle for effective shot deblending. Conventional deblending methods require determining the specific excitation time of each shot, and based on this, performing operations such as pseudo deblending, channel set conversion, and denoising. This not only requires high accuracy of the excitation time, but also is a complicated operation that requires denoising each shot separately, which is computationally huge. We designed a multi-output U-shaped Net Transformer (UNetr)based on the principles of imaging. By utilizing a transformer, which is more sensitive to positional information, as an encoder, this network can distinguish the waveform characteristics of different single shots and separate the blended data directly in the common shot channel set. After testing, the method is more capable for coherent signals and more effective for deblending of overlapping shots. Without relying on time coding, the method skips the complex processing flow. The processing efficiency is improved and the deblending effect is significant.

Abstract The classification of seismic event types is a fundamental prerequisite for reliable monitoring of seismic activity and for scientific assessment of earthquake hazards. However, efficiently and accurately distinguishing between natural earthquakes and artificial blast events based on seismic observation data from stations remains a key research topic with both challenges and value in the fields of seismology and signal processing. To address this issue, this article proposes a seismic event classification and prediction algorithm based on a deep learning network. The algorithm adopts a dual-branch joint decision-making strategy, taking single-channel waveform data with simple preprocessing as an input. During the model inference process, feature extraction and prediction are performed separately on the waveform data and its corresponding time–frequency matrix. In addition, a weight learning subnetwork is designed to adaptively adjust the weight proportion of the prediction results from the two branches in the overall decision-making, thereby ensuring the stability and accuracy of the classification and prediction results. Experimental test results show that the application process of the algorithm is simple, without relying on extensive manual parameter tuning, and it exhibits good generalization performance in seismic data sets from different observation instruments and different regions.

This study investigates how solar and geomagnetic driver selection affects 24 h-ahead global ionospheric vertical total electron content (VTEC) prediction under different geomagnetic conditions. A four-step feature selection strategy involving importance evaluation, redundancy elimination, physical interpretability prioritization, and performance validation was developed to identify five key drivers from candidate solar and geomagnetic factors. Using global ionospheric maps provided by the Center for Orbit Determination in Europe (CODE) from 2014 to 2018, a non-overlapping 90-day temporal block scheme was adopted to reduce the risk of temporal information leakage. Six ablation experiments were conducted to compare the predictive performance of different driver combinations. The results show that the full-factor configuration selected by the proposed strategy achieved the most favorable overall performance among the tested combinations, although the global-average improvement relative to the baseline remained modest. The optimal driver combination varied with geomagnetic disturbance level, and the contribution of external drivers showed clear latitudinal dependence. In addition, the full-factor configuration yielded a more balanced global error distribution and was associated with slower error accumulation over the 24 h horizon. These findings suggest that physically guided driver selection is useful for constructing more physically meaningful driver combinations and for improving long-horizon prediction stability within a unified ConvLSTM-based framework.

The quality of the wavelet significantly affects the accuracy of full-waveform inversion (FWI). In practice, seismic data are often contaminated by non-ideal wavelets with complex side lobes and ambiguous dominant frequencies. Directly employing such data in conventional FWI workflows, which typically assume an ideal Ricker wavelet, can induce severe cycle-skipping problems and lead to inversion failure. Existing remedies, such as sophisticated source wavelet estimation or the design of wavelet-independent objective functions, often suffer from instability in estimation or the loss of waveform details. To address this issue at its root from a data-driven perspective, this study proposes a novel deep learning-based approach. We introduce a multi-scale Swin Transformer transposed U-Net (MSTU) network to intelligently transform seismic records containing unknown, non-ideal wavelets into universal records composed solely of ideal Ricker wavelets. Through end-to-end learning, our method establishes a direct nonlinear mapping from the complex wavelet domain to the universal wavelet domain, effectively stripping the inherent wavelet ambiguity from the raw data. Numerical examples demonstrate that data universal by the MSTU network can be directly fed into standard FWI routines, yielding superior inversion results. For data originally generated with Klauder or Ormsby wavelets, the inversion results obtained after our standardization process significantly outperform those from direct inversion using the incorrect wavelet or even using the original non-ideal wavelet itself. Moreover, the accuracy approaches the theoretical upper limit achievable by inversion with the ideal Ricker wavelet as the ground truth. Requiring no prior wavelet estimation or modification to the core inversion algorithm, our method serves as an efficient, robust, and “plug-and-play” data preprocessing tool, offering a highly practical solution to enhance the reliability of FWI for field seismic data.

The location of passive seismic events is essential for seismic activity studies. Deep learning models have shown promising results in localizing source locations from seismic waveforms. However, these methods suffer from deficiencies in generalization and interpretability. These well-trained models may only be well adapted to the region where the training data are located. Here, we propose a passive seismic source localization method based on knowledge-augmented deep learning (KADL). The proposed method integrates the scientific knowledge of time-reversal imaging into the deep learning model at the data level and architecture level, respectively. We design a physics guidance module for the effective integration of knowledge and data. The introduction of scientific knowledge provides guidance and constraints for the training of the model, helping it to focus more efficiently on key features related to passive seismic source localization. The experiment results show that the integration of scientific knowledge and deep learning can achieve better generalization. Especially in the application scenarios across different datasets, the accuracy of source location is significantly improved. The KADL model has the potential to be generalized to different regions. We also use explainable artificial intelligence to analyze the importance of data and knowledge, providing valuable insights into the decision-making process of the KADL model.

Accurate inversion of iron oxide abundance is crucial for mineral exploration and geological assessment. However, traditional spectral inversion methods are susceptible to interferences from factors such as particle size, topography, and mixed pixels, which limit prediction accuracy. To address this issue, this paper proposes a quantitative inversion method for iron oxide abundance using hyperspectral remote sensing, combining Wavelet Packet Analysis (WPA) and Partial Least Squares Regression (PLSR). This method utilizes the optimal wavelet packet basis function to perform multi-scale decomposition on laboratory mineral spectra, separating approximation signals and detail signals, and combining with PLSR to establish a high-precision abundance prediction model. Subsequently, the optimal model is applied to GaoFen-5 (GF-5) AHSI hyperspectral satellite data. The results show that the spectral detail signals can not only effectively amplify the absorption features of iron oxide (sensitive bands at $400-1200 \text{nm}$), significantly improving the fitting accuracy, but also effectively separate and remove the systematic stripe noise in the hyperspectral images during the image reconstruction phase. By effectively isolating high-frequency spectral features, this combined model successfully mitigates the interferences of topography and particle size effects, achieves high-precision spatial distribution mapping of mineralization zones in the study area, and provides reliable technical support for the delineation of prospecting targets.

This study introduces a novel deep learning framework for the integrated processing of active and passive seismic data, enabling concurrent seismic inversion and imaging. By leveraging the time-frequency complementarity of multisource data and the mutual constraints between inversion and imaging, the proposed approach achieves high-resolution subsurface imaging and accurate velocity model inversion. Numerical experiments on synthetic datasets demonstrate the model’s ability to recover subsurface structures with high precision, even in the presence of noise and data incompleteness, underscoring its robustness.

Full waveform inversion (FWI) using passive seismic data offers potential advantages for obtaining deep subsurface structural information. However, its application depends on a reliable background model and accurate source locations. Here, we introduce a joint inversion framework that uses virtual and actual passive seismic data to achieve velocity modeling and source localization. This method first retrieves the broadband virtual reflection responses at the surface by seismic interferometry. Then, the virtual responses are used as the basis for multiscale FWI to recover the subsurface velocity structures without knowledge of the actual passive source locations. Finally, using the inversion result as the initial model, alternately perform reverse-time wavefield scanning and FWI with actual passive-source data, enabling simultaneous inversion for source locations and velocity information and compensating for deep illumination. Numerical examples under various conditions of inhomogeneous passive-source distribution and signal-to-noise ratio (SNR) show that the proposed method can robustly recover the subsurface velocity structures using only passive seismic data, even when the initial model is rough, while accurately locating passive seismic events.

In recent years, the attention paid to the viscoacoustic wave equation with decoupled fractional Laplacian (DFL) operators has increased remarkably. This is due to the equation's unique decoupling property and its ability to accurately characterize the quality factor. The DFL viscoacoustic equation is typically computed numerically employing the finite difference (FD) and Fourier pseudo-spectral (PS) methods with respect to the derivatives of time and space, respectively. Although the FD-PS method provides spectral accuracy in the spatial domain, its time accuracy is limited to second order. In scenarios with a larger time sampling interval, the FD-PS method might encounter significant time dispersion. Although reducing the time sampling interval can alleviate this dispersion, it significantly boosts computational cost. To correct the time error introduced by time dispersion, we propose a hybrid rapid expansion method (HREM) to rectify the time error in DFL viscoacoustic equation. This method mitigates the time error by compensating for the dispersion-dominated term that exerts a considerable influence on the time error. HREM has been validated to effectively correct the time error through theoretical analysis and numerical experiments. Compared to the traditional FD-PS method, the stability condition of HREM is looser, thereby conferring enhanced flexibility in the selection of sampling parameters. Furthermore, HREM allows for the accurate estimation of the wavefield with larger time steps. The application of a three-dimensional overthrust model illustrates its effectiveness for large-scale seismic modelling.

Seismic interferometry using ambient noise provides an effective approach for subsurface imaging through reconstructing passive virtual source (PVS) responses. Traditional crosscorrelation (CC) seismic interferometry relies on a uniform dense distribution of passive sources in the subsurface, which is often challenging in practice. The multidimensional deconvolution method (MDD) alleviates reliance on passive-source distribution, but requires wavefield decomposition of the original data. This is difficult to accurately achieve for uncorrelated noise sources, leading to the existence of non-physical artifacts in the reconstructed PVS data. To address this issue, this study proposes a method to improve the accuracy of PVS data reconstruction using an enhanced U-Net. This data-driven approach circumvents the challenge of noise wavefield decomposition encountered in the traditional MDD. By integrating a feature fusion module into U-Net, multi-scale sampling information is leveraged to improve the network’s ability to capture detailed PVS data features. The combination of active-source data constraints and the modified MDD further optimizes PVS data retrieval during training. Numerical tests show that the proposed method effectively recovers waveform information in PVS retrieval records with non-ideally distributed sources, suppressing coherent noise and false events. The reconstructed recordings have a clear advantage in the reverse time migration (RTM) imaging results, with strong generalization performance across various velocity models.

Simultaneous-source acquisition technology enables seismic data to overlap,overcoming the time and spatial constraints of conventional acquisition methods and significantly enhancing seismic data acquisition efficiency.However,this approach introduces substantial aliasing noise,increasing the complexity of subsequent data processing.Since aliasing noise exhibits a discrete random distribution in non-common shot domains,separating blended seismic data can be treated as a noise removal process.This study employs deep learning techniques to achieve the separation of blended seismic data.To address issues such as partial feature loss and insufficient feature fusion accuracy in the original U-Net's down-sampling process,the U-Net++architecture is adopted.By incorporating dense connections and a deep supervision mechanism,the model's training process is optimized.The trained network is then applied to separate mixed seismic data.Experimental results on simulated and real data demonstrate that the U-Net++architecture outperforms the original U-Net,achieving superior separation performance with notable improvements in signal-to-noise ratio and data fidelity.

Full waveform inversion (FWI) seeks a subsurface parameter model that optimally matches the true state by minimizing the differences between synthetic and observed data. However, when starting from a rough initial model, FWI is often limited by the weak low-frequency energy of the observed data and the difficulty of matching surface-related multiples (SRMs), especially when the source wavelet is not readily available. Source wavelet errors also affect the general inversion result. We propose a multiscale virtual wavefield waveform inversion (VWWI) based on multidimensional interferometric retrieval (MDIR) to mitigate these challenges. We use MDIR to retrieve the virtual response from the up- and down-going wavefields separated from the original data and infer the velocity using the virtual response instead of the original data. MDIR integrates the source functions using multidimensional cross correlation (MDCC) and then suppresses the source imprints from the original data through multidimensional deconvolution (MDD). The retrieved virtual responses have a broader bandwidth and are dominated by primary reflection events. It addresses simultaneously three major challenges that FWI faces through a one-time data retrieval. Assigning self-setting source functions with different dominant frequencies to the virtual response allows the extraction of virtual observed data to different frequency bands for multiscale velocity inversion. Considering the possible amplitude distortion and the computational cost, we propose the hybrid source cross-correlation objective function adapted to VWWI. Numerical examples of well-known models representing weak and strong scattering media show that the proposed VWWI method can stably achieve wide-scale velocity modeling from macroscopic background to delicate structures.

Full waveform inversion (FWI) comprehensively utilizes phase and amplitude information of seismic waves to obtain high-resolution subsurface medium parameter models, applicable to both active-source and passive-source seismic data. Passive-source seismic exploration, using natural earthquakes or ambient noise, reduces costs and environmental impact, with growing marine applications in recent years. Its rich low-frequency content makes passive-source FWI (PSFWI) a key research focus. However, PSFWI inversion quality relies heavily on accurate virtual source reconstruction. While multi-dimensional deconvolution (MDD) can handle uneven source distributions, it struggles with irregular receiver sampling. We propose a robust MDD method based on multi-domain stepwise interpolation to improve reconstruction under non-ideal source and sampling conditions. This approach, validated via an adaptive PSFWI strategy, exploits MDD’s insensitivity to source distribution and incorporates normalized correlation objective functions to reduce amplitude errors. Numerical tests on marine and complex scattering models demonstrate stable and accurate velocity inversion, even in challenging acquisition environments.

Due to the limitations of seismic exploration instruments and the impact of the high frequencies absorption by the earth layers during subsurface propagation of seismic waves, recorded seismic data usually lack high and low frequency information that is needed to accurately image geological structures. Traditional methods face challenges such as limitations of model assumptions and poor adaptability to complex geological conditions. Therefore, this paper proposes a deep learning method that introduces the attention mechanism and Bi-directional gated recurrent unit (BiGRU) into the Transformer neural network. This approach can simultaneously capture both global and local characteristics of time series data, establish mappings between different frequency bands, and achieve information compensation and frequency extension. The results show that the BiGRU-Extended Transformer network is capable of compensating and extending the synthetic seismic data sets with the limited frequency band. It has certain generalization capabilities and stability and can effectively handle various problems in the data reconstruction process, which is better than traditional methods.

In passive source seismic surveys, signal continuity and signal-to-noise ratios have always tended to be low. On the one hand, since passive-source seismic surveys are often used for large-scale illumination of subsurface formations, the distances between receivers and sampling point intervals tend to be large. On the other hand, interference from coherent noise and spurious in-phase axes is unavoidable in passive source reconstruction recordings because of the signal originating from noise in the subsurface. All these problems lead to the continuity and signal-to-noise ratio of the virtual shot reconstructed from passive source seismic surveys are not guaranteed, which affects further processing and seriously limits the application of passive source seismic surveys. The traditional interpolation reconstruction methods cannot take noise suppression into account, or require additional operations to achieve both interpolation reconstruction and denoising. Based on this, this paper utilizes the powerful data processing ability of convolutional neural networks to design a global multi-scale fusion residual shrinkage network (GMF-RS) to solve the above passive source seismic exploration problem. It is tested that the trained network not only eliminates coherent noise and false events, but also improves the continuity in horizontal and vertical directions, enhances and extracts the effective signals, and provides better virtual shot records for subsequent seismic data processing. In addition, we designed a dual-input network and introduced active source seismic records as a complement to the passive source virtual seismic records, so that the processed waveforms can show better details.

In passive seismic exploration, the number and location of underground sources are very random, and there may be few passive sources or an uneven spatial distribution. The random distribution of seismic sources can cause the virtual shot recordings to produce artifacts and coherent noise. These artifacts and coherent noise interfere with the valid information in the virtual shot record, making the virtual shot record a poorer presentation of subsurface information. In this paper, we utilize the powerful learning and data processing abilities of convolutional neural networks to process virtual shot recordings of sources in undesirable situations. We add an adaptive attention mechanism to the network so that it can automatically lock the positions that need special attention and processing in the virtual shot records. After testing, the trained network can eliminate coherent noise and artifacts and restore real reflected waves. Protecting valid signals means restoring valid signals with waveform anomalies to a reasonable shape.

The pivotal role of seismic velocity inversion in oil and gas exploration and geological research has been widely acknowledged. However, conventional methods face challenges such as strong reliance on initial models and high computational costs. Based on the mode of seismic event generation, seismic data can be classified into active seismic data and passive seismic data, which collectively constitute the multisource data discussed in this article. Velocity inversion based on deep learning primarily relies on active seismic data, training neural networks to learn the mapping between seismic records and subsurface velocities. In contrast, signals in passive seismic data typically originate from noise at certain depths within the Earth, encompassing valuable information about deep subsurface structures that is crucial for velocity inversion, thus presenting a potential complement to active seismic data. This study proposes a seismic velocity inversion method that combines active and passive seismic data, utilizing deep learning techniques to adaptively integrate data from both sources, enabling joint inversion. The proposed neural network architecture combines transformer and convolutional neural network (CNN), enhancing the accuracy and robustness of velocity inversion.

Microseismic source localization methods with deep learning can directly predict the source location from recorded microseismic data, showing remarkably high accuracy and efficiency. Two main categories of deep learning-based localization methods are coordinate prediction methods and heatmap prediction methods. Coordinate prediction methods provide only a source coordinate and generally do not provide a measure of confidence in the source location. Heatmap prediction methods require the assumption that the microseismic source is located on a grid point. Thus, they tend to provide lower resolution information and localization results may lose precision. This study reviews and compares previous methods for locating the source based on deep learning. To address the limitations of existing methods, we devise a network fusing a convolutional neural network and a Transformer to locate microseismic sources. We first introduce the multi-modal heatmap combining the Gaussian heatmap and the offset coefficient map to represent the source location. The offset coefficients are utilized to correct the source locations predicted by the Gaussian heatmap so that the source is no longer confined to the grid point. We then propose a fusion network to accurately estimate the source location. A gated multi-scale feature fusion module is developed to efficiently fuse features from different branches. Experiments on synthetic and field data demonstrate that the proposed method yields highly accurate localization results. A comprehensive comparison of coordinate prediction method and heatmap prediction methods with our proposed method demonstrates that the proposed method outperforms the other methods.