Node seismometer signals are often contaminated with substantial environmental noise due to the complex conditions encountered in geophysical exploration and seismic monitoring. This necessitates high precision in the pre-processing of ground motion signals, as inadequate processing may compromise subsequent operations, such as P-wave first-arrival time extraction, peak energy calculation, ground motion period determination, and magnitude estimation. To obtain more authentic seismic waveforms, a node seismometer signal denoising model is proposed. This model integrates grey relational analysis (GRA) with improved complete ensemble empirical modal decomposition adaptive noise (ICEEMDAN). This method first decomposes the noisy signal using ICEEMDAN to obtain multiple intrinsic mode functions (IMFs), which are then sequentially arranged and labeled. Subsequently, for each IMF, the correlation coefficient, mutual information, R ², adjusted R ², Jensen–Shannon divergence, cosine similarity, root mean squared error, mean absolute error, mean absolute percentage error, and sample entropy were calculated, forming an evaluation matrix for assessing the reliability of all IMFs. Finally, using GRA, the correlation coefficients and degrees of association between each evaluation metric and different IMF components were calculated. The IMF components were ranked based on their association degrees to determine the relative effectiveness of their signal components. Linear reconstruction on the top-ranked IMF components was performed to complete the signal denoising process proposed. Experiments on denoising simulated seismic signals, recorded seismic event signals, and recorded ground motion signals all demonstrate that the GRA–ICEEMDAN model outperforms classical denoising methods. The comprehensive denoising scores for the three experiments were 100, 98.0180, and 93.9056, respectively, with signal-to-noise ratio improvements reaching 24.0049 dB, 20.8926 dB, and 16.3523 dB, respectively. The model effectively distinguishes noise components from effective components, with minimal reconstruction errors

The development of coalbed methane (CBM) relies on high-precision reservoir prediction and lithological inversion. Seismic amplitude variation with offset (AVO) technology is an important tool for fine-scale reservoir characterization. However, the seismic AVO response of CBM reservoirs is complex and is affected by seismic rock physics parameters at different scales. Microscopically, aligned fractures in CBM reservoirs produce complex anisotropy due to formation inclination. At the macroscopic scale, the thickness of CBM reservoirs within seismic frequency bands is comparable to the seismic wavelength and should therefore be treated as a layered medium. In addition, pore fluid significantly affects seismic wave propagation. Consequently, determining the azimuthal AVO response of CBM reservoirs in relation to seismic rock physics parameters at different scales can support high-precision reservoir prediction and lithological inversion. In this study, the primary–primary wave reflection coefficient for a two-phase layered medium was derived using Biot’s theory. Using this model, the response characteristics of the reflection coefficients with respect to seismic azimuth, aligned fracture parameters, reservoir thickness, and seismic main frequency were analyzed. A rotated staggered–grid finite–difference algorithm was employed to simulate wavefield characteristics separately for coal seams and surrounding strata. Seismic attributes were then used to characterize the seismic AVO response. The Z-direction seismic amplitude attributes and reflection coefficients showed similar trends in their responses to seismic rock-physical parameters. This study contributes to establishing a more precise seismic AVO response framework to support CBM reservoir prediction and high-quality lithological inversion.

Full-waveform inversion (FWI) imaging is a high-resolution seismic imaging technique that directly produces subsurface images by inverting the full recorded wavefield. However, its reliability is often limited by numerical dispersion errors arising from finite-difference (FD) forward modeling. One key approach for reducing dispersion is to optimize the FD coefficients using an optimization algorithm. However, conventional methods for optimizing FD weights focus only on reducing spatial dispersion, which can weaken numerical stability, especially when using large time steps (i.e., high Courant–Friedrichs–Lewy [CFL] numbers). To address this issue, we introduce a new optimization approach that improves both simulation accuracy and stability. The proposed method combines error functions from both the time–space domain and the spatial domain into a single adaptive objective function. A dynamic weighting factor, which depends on the CFL number, facilitates a trade-off between accuracy and stability of the optimal FD weights. We also use the seismic wavelet spectrum as prior information to constrain the optimization. The optimization problem is solved by the least-squares method. In the theoretical test, the proposed weights significantly reduce wavefield simulation errors across a wide range of wavenumbers, with a higher CFL number than conventional approaches. When applied to FWI, this method reduces phase distortion and local minima in the objective function. In a test using the Marmousi model at 40 Hz, our approach produced clear and continuous deep structures, closely matching results from dispersion-free benchmarks. In contrast, conventional methods failed due to severe dispersion. This work provides a more robust numerical foundation for high-frequency FWI imaging by improving both accuracy and stability.

Global warming is causing permafrost thawing, which can lead to permafrost collapse. Assessing the stability of permafrost against this collapse risk is essential. One effective approach to evaluating permafrost stability is to use the elastic modulus. Since the elastic modulus varies with the ice content in the pore spaces, analyzing its relationship with porosity is crucial for understanding permafrost stability. Previous studies have used rock cores to analyze the relationship between elastic modulus and porosity. However, the analysis of elastic modulus relative to various porosities in permafrost is limited. To overcome these limitations, porosity was controlled using 3D printing, and a permafrost analog model with seismic velocities and porosities similar to those of natural permafrost was fabricated and assumed as permafrost. To analyze elastic modulus as a function of porosity in permafrost, seismic velocities were measured through seismic physical modeling experiments, and the elastic modulus was estimated from the measured seismic velocities. Analysis of the relationship between elastic modulus and porosity revealed that the elastic modulus decreased after increasing to a specific porosity in the permafrost simulation model. These findings provide a quantitative basis for evaluating the stability of infrastructure on permafrost by demonstrating the significant effect of ice expansion on mechanical properties. Furthermore, this study validates the 3D printing approach as an effective tool to overcome the limitations of natural rock cores for systematic permafrost research.

Deep coalbed methane exploration in the Ordos Basin holds significant potential, yet the frequent development of coal gangue severely hinders productivity and complicates horizontal well deployment. Accurately characterizing the spatial distribution of these thin coal gangue layers remains a critical challenge, particularly when constrained by limited seismic resolution and sparse well control. To address this challenge, we propose a novel paleogeomorphology-constrained stochastic seismic inversion workflow applied to the Benxi Formation in the YL area. This approach integrates petrophysical analysis with paleogeomorphic restoration, identifying natural gamma as the most sensitive lithological indicator. Under hierarchical geomorphic constraints, geostatistical simulation was utilized to predict the three-dimensional spatial probability of coal gangue occurrence. Quantitative validation demonstrates a robust correlation between coal gangue thickness and natural gamma response, with the inversion results achieving an accuracy of 81.82% in blind well validation. Spatially, gangue development is controlled by paleotopography, with higher probabilities concentrated in paleo-highs and slopes associated with stronger hydrodynamic conditions, while paleo-depressions exhibit superior coal continuity. This study not only overcomes the resolution limitations of traditional inversion in sparse-well areas but also provides a rigorous quantitative geological basis for sweet-spot identification and trajectory optimization in deep coalbed methane development.

Seismic monitoring in extreme environments, such as arid regions adjacent to the Alxa Desert, faces significant challenges due to complex noise interference from dust storms, high wind noise, and thermal variations. This paper presents TFDenoiser-Edge, a novel hybrid deep learning framework that combines convolutional neural networks (CNN) and Transformer architectures for real-time seismic signal denoising on resource-constrained edge devices. The proposed model employs a U-Net encoder–decoder structure with Transformer modules for global feature modeling in the time-frequency domain. To enable deployment on edge neural processing units (NPUs) with limited memory (≤512 MB), we introduced a mixed-precision quantization strategy that applies INT8 quantization to CNN layers while maintaining BF16 precision for Transformer layers, achieving 3.6× model compression with only 0.3 dB signal-to-noise ratio (SNR) loss. Additionally, a block-wise computation approach reduces peak memory consumption from 86 MB to 7.8 MB. Extensive experiments on Gansu seismic data demonstrated that TFDenoiser-Edge achieved an average SNR improvement of 8.5 dB, with P-wave and S-wave detection rates increasing from 65% to 91% and 52% to 85%, respectively. The model achieved real-time inference with 68 ms latency on edge NPUs while consuming less than 5 W of power, making it suitable for autonomous seismic monitoring in arid and desert regions. The proposed framework demonstrates potential generalizability to other extreme environments through transfer learning with minimal fine-tuning.

Coal-bearing strata hold significant potential for critical metal resources. This study investigates their enrichment mechanisms in the eastern Ordos Basin using an integrated geophysical-geochemical approach. We combine seismic impedance inversion, tectonic evolution, and log facies interpretation with geochemical data (total organic carbon, sulfur, vitrinite reflectance, trace/rare earth elements) to reconstruct the sedimentary-tectonic-hydrogeological evolution. Results show notable enrichment of zirconium, chromium (Cr), nickel, and hafnium. Seismic facies indicate that deltaic distributary channels and tidal sandbars were primary conduits for metal-bearing fluids. The Taiyuan Formation was deposited in a warm-humid, freshwater-influenced tidal delta-barrier island system under reducing conditions, enabling initial metal sequestration via sulfidation and organic complexation. Reconstruction of tectonic evolution confirms that Yanshanian tectono-thermal events associated with the Zijinshan pluton generated fault networks that channeled hydrothermal fluids, driving secondary metal enrichment. Groundwater circulation, controlled by sedimentary facies and fractures, further regulates metal remobilization. A focused study on Cr delineates a three-stage enrichment model: source weathering, sedimentary reduction, and tectonic-hydrothermal activation. This work establishes a structure–sedimentation–groundwater coupled joint ore-controlling model, underscoring the value of integrated geophysical and geochemical methods for exploring critical metals in coal-bearing strata.

Traditional full waveform inversion objective functions typically rely on discrepancies between waveforms, making full waveform inversion highly dependent on the initial model and the low-frequency components in the seismic data. The instantaneous phase of the seismic signal reflects the kinematic information of seismic waves and has a more linear relationship with the subsurface velocity structure, aiding in the retrieval of low-wavenumber components of velocity models. Conversely, waveform information is instrumental in achieving high-resolution inversion results while maintaining algorithmic stability. In addition, automatic differentiation provides an efficient and accurate mechanism for computing gradients by systematically applying the chain rule across the computational graph, yielding reliable derivative information for optimization. To balance the contributions of waveform and phase information in velocity inversion, within the framework of automatic differentiation, we integrate instantaneous phase and waveform information to propose an automatic differentiation-based weighted instantaneous phase inversion method. Numerical tests using the Marmousi model and the Overthrust model demonstrate that the proposed method achieves more accurate velocity inversion results.

Accurate velocity models that include long-wavelength components are required for precise subsurface structure imaging and estimation of geophysical properties. To successfully build the long-wavelength velocity using full waveform inversion (FWI), sufficient offset and low-frequency components are necessary. However, due to the limited acquisition conditions, it is difficult to operate long offset and low-frequency sources in coastal shallow marine. Short offset makes the data dominated by reflections, and FWI intensively updates the surface boundaries. Although reflection full waveform inversion (RFWI) has been proposed to reconstruct long-wavelength velocity models using reflection data, it suffers from cycle skipping when low-frequency component is insufficient. To overcome these limitations, we propose a direct envelope-based RFWI (DE-RFWI) that incorporates the direct envelope into the RFWI. By employing envelope-based energy information from reflection data, DE-RFWI facilitates the reconstruction of long-wavelength velocity. The proposed method employs the Hilbert transform-based implicit gradient decomposition technique to address additional computational cost. To verify the proposed method, DE-RFWI was applied to synthetic test with a shallow marine condition and field data acquired in Yeongil Bay, South Korea. The inversion results for field data were evaluated by analyzing arrival-time alignment in envelope domain. The results demonstrate that DE-RFWI can reliably reconstruct long-wavelength velocity models in shallow marine seismic data and improve reflector continuity and resolution in the reverse time migration imaging.

Amplitude variation with incident angle (AVA) inversion for partial angle stack seismic data is an extension of acoustic impedance inversion. It adopts the P-wave reflection coefficient to link seismic amplitude information with elastic parameters and incident angle. This method generalizes conventional acoustic impedance inversion to pre-stack seismic data, enabling the effective inversion of multiple elastic parameters by utilizing seismic data acquired at different incident angles, such as angle gather data or partial angle stack data. This study addresses the limitations of commonly used AVA inversion with Cauchy regularization. Due to the non-linearity of Cauchy regularization, the iterative re-weighted least squares algorithm is extensively employed. It has been shown that the accuracy of inversion results of this non-linear inversion algorithm is highly dependent on the initial solution. When geological conditions are complex, the initial solution must be close to the optimal solution to ensure accurate inversion results. To address this challenge, this study proposes combining the ensemble smoother with multiple data assimilation (ES-MDA) and AVA inversion with Cauchy sparse regularization. Specifically, the posterior mean of ES-MDA is used as the initial solution for AVA inversion with Cauchy sparse regularization. ES-MDA is a stochastic method that solves inverse problems by iteratively updating an ensemble of model realizations, yielding a solution that closely approximates the optimal solution. Practical application indicates that, compared with AVA inversion with Cauchy sparse regularization using a standard initial solution, the proposed method achieves improved accuracy in estimated elastic parameters. The research findings offer new technical approaches for seismic prediction with AVA inversion in complex reservoirs.

The Linxing–Shenfu gas field, a key block for China National Offshore Oil Corporation’s onshore unconventional oil and gas exploration, is characterized by complex geology that poses dual challenges to the efficiency and accuracy of traditional seismic interpretation methods. This study presents a systematic review of the application progress of machine learning, particularly deep learning, in seismic interpretation within this block since 2018. To address the specific geological characteristics and exploration needs, we developed a comprehensive intelligent interpretation workflow. This workflow integrates intelligent horizon and fault interpretation, deep clustering for seismic facies analysis, and automated identification of special geological bodies (e.g., Zijinshan igneous rock mass), enabling the accurate reconstruction of the stratigraphic framework. Furthermore, leveraging deep learning models, we achieved direct prediction of lithology, physical properties (e.g., porosity, permeability), and gas-bearing parameters, culminating in the comprehensive characterization of geological “sweet spots.” Practical applications demonstrate that this intelligent interpretation workflow not only significantly enhances interpretation efficiency but also provides distinct advantages for overcoming the bottlenecks of traditional theoretical methods, such as handling low signal-to-noise ratio data, identifying thin interbeds, and predicting “sweet spots.” This review provides robust support for efficient exploration and development decision-making in the Linxing–Shenfu Block.

Privacy protection and multimodal fusion represent significant challenges in the context of cross-institutional geological data collaboration. To address these challenges, this paper proposes a collaborative XGBoost-Deep Residual model (FedXGB-ResNet) within a federated learning framework to achieve high-accuracy porosity prediction. Through a heterogeneous federated integration architecture, the framework combines XGBoost’s efficient modeling of structured well-logging parameters (with a feature importance gain of 41.2%) with ResNet’s deep extraction of spatiotemporal features from seismic images (87% overlap in activation maps between training and validation sets). The model achieved R² scores of 0.87 and 0.83 on North Sea and Bakken oilfield datasets, respectively, representing improvements of 12.7% and 9.3% over the traditional FedAvg-XGBoost baseline. The innovative dual privacy protection mechanisms—gradient obfuscation and Paillier encryption—suppressed the membership inference attack success rate to 13.7% and reduced gradient similarity to 0.19 at a noise scale of 1.5 (privacy budget = 0.75), while increasing communication time by only 23%. A dynamic feature distillation mechanism adaptively fuses multimodal features through gated attention units, narrowing the F1-score gap between high-porosity and low-porosity identification to 5%. Experiments demonstrated that the framework reduces the risk of privacy leakage by 84.7% while retaining 93.4% of the performance of a centralized model, offering a balanced solution for collaborative cross-domain geological data analysis in terms of accuracy, privacy, and efficiency.

The application of deep learning to seismic fault interpretation is often constrained by computational costs. To address this, we propose a novel strategy that decouples the context window size (the spatial range of seismic observations) from the input resolution (the actual matrix dimensions fed into the model), systematically investigating their combined impact on computational efficiency and interpretation accuracy. Using field-acquired seismic data from two distinct coal mines, we trained a lightweight two-dimensional convolutional neural network (CNN) on samples extracted with varying context windows (8 × 8 to 64 × 64 pixels), which are then uniformly resized to a fixed low resolution of 8 × 8 pixels. Our results demonstrate that enlarging the context window consistently improved model performance, with the 64 × 64 window achieving the highest precision (99.46%) and fault continuity, even after downscaling. In contrast, a combined multi-scale training set did not outperform the best single-window model, indicating that effective multi-scale fusion requires more advanced architectural designs. Our workflow highlights that contextual information remains crucial for feature learning despite input standardization, and offers an efficient paradigm: large context window + small fixed input + lightweight network, that maintains high accuracy while significantly reducing computational costs. This approach provides a practical pathway for deploying deep learning models in resource-limited geophysical applications.

Fault identification is a critical task in 3D seismic interpretation, directly influencing the efficiency and accuracy of reservoir characterization and drilling decisions. However, traditional methods rely heavily on manual experience and high-quality annotated data, making it difficult to adapt to the demands of processing massive amounts of seismic data. To address this, we introduce a novel self-supervised learning (SSL) paradigm specifically designed for geological feature learning, which for the first time unifies masked voxel reconstruction, 3D patch contrastive learning, and multimodal attribute joint contrast into a coherent multi-task pre-training framework. This framework is uniquely tailored to leverage the spatial continuity and physical attributes of seismic data, enabling the model to learn transferable, geologically meaningful representations without any manual labels. It leverages unlabeled seismic data to learn geologically meaningful feature representations and leverages a transfer learning mechanism to achieve high-precision fault identification with small sample sizes. Experiments on multiple public and field-measured datasets, including F3 and SEAM Phase I, demonstrate that this method achieves key metrics such as intersection over union (IoU) and F1-scores of 0.76 and 0.87, respectively, significantly outperforming traditional attribute analysis (IoU = 0.49) and supervised deep learning models (IoU = 0.72). Furthermore, the method remains robust in areas with low signal-to-noise ratios (average confidence > 85%) and consistently estimates fault strike and dip (average absolute error ≤ 3.5°). This research provides an effective solution for reducing reliance on manual annotation and improving the reliability of automated fault interpretation in complex tectonic areas.

Accurate estimation of Poisson’s ratio is essential for the characterization of coal-bearing gas reservoirs, particularly in thin-bed and low signal-to-noise ratio environments where conventional elastic impedance (EI) inversion suffers from wavelet interference, limited resolution, and reliance on linearization assumptions. To address these limitations, we develop a physics-guided two-branch convolutional neural network (TB-CNN) that directly predicts Poisson’s ratio by jointly integrating EI-inverted P-wave velocity, S-wave velocity, and density with small-, medium-, and large-angle seismic stacks. The first branch provides geologically consistent, physics-informed background trends, while the second branch captures thin-bed-sensitive reflectivity features and amplitude tuning effects. The fused latent representation is explicitly regularized using empirical rock-physics relationships to ensure physical plausibility and enhanced generalization. Field validation on the 8# coal seam of the Ordos Basin demonstrates that the proposed TB-CNN improves vertical resolution, sharpens seam boundary delineation, and better preserves thickness variations compared with EI inversion and a single-branch CNN. Near-well comparisons show higher correlation with log-derived Poisson’s ratio, while lateral slices reveal improved continuity and thin-layer detectability. These results confirm that combining physics-guided stability with data-driven resolution provides a robust and interpretable framework for Poisson’s ratio inversion in thin coal seams and holds promise for broader applications in unconventional gas reservoir prediction.

As hydrocarbon exploration advances toward deep and complex reservoirs, the identification accuracy of traditional time–frequency (TF) analysis is constrained by strongly heterogeneous geological conditions. Concurrently, while deep learning has shown great potential, mainstream supervised models commonly face the dual challenges of scarce labeled samples and the lack of interpretability in their black-box decision-making processes. To address these challenges, this study proposes an innovative, intelligent prediction framework integrating high-precision TF analysis, manifold-regularized semi-supervised generative adversarial networks, and SHapley Additive exPlanations (SHAP) for interpretability. First, the Fourier-based synchrosqueezing transform was employed to extract two-dimensional TF features with superior energy concentration, effectively overcoming the resolution limits imposed by the Heisenberg uncertainty principle. Subsequently, the manifold-regularized semi-supervised generative adversarial network was developed. By incorporating manifold regularization constraints, the discriminator captures the intrinsic topological structure of large-scale unlabeled samples, effectively leveraging data geometry to significantly enhance generalization capability under sparse-label conditions. Finally, the SHAP method was utilized to conduct a post hoc interpretation. Experimental results on the Marmousi II model demonstrate a remarkable testing accuracy of 98.4%. In a real-world application to deep marine reservoirs in the Sichuan Basin, the framework achieved an 85.0% testing accuracy using only 5% labeled samples. Compared to baseline models, the semi-supervised strategy and manifold regularization contributed accuracy gains of 18.8% and 5.0%, respectively. SHAP analysis further confirms the model’s adaptive capability to extract geophysical features, enabling it to deconstruct the tuning-effect patterns in synthetic data and the low-frequency enhancement/high-frequency attenuation patterns in real data, respectively. This validation of geophysical consistency provides a theoretical foundation for the application of artificial intelligence in complex hydrocarbon exploration.

Anisotropy is a prevalent characteristic of the Earth’s subsurface, giving rise to more complex wavefield behavior than in purely acoustic media. Seismic data acquired in such anisotropic environments pose significant challenges for conventional acoustic imaging methods that rely on isotropic assumptions and acoustic approximations, often leading to inaccurate imaging of complex geologic structures. To address these challenges, we propose a wavefield depth extrapolation method grounded in the two-way acoustic anisotropic wave equation, specifically tailored for imaging in vertically transversely isotropic (VTI) media. Theoretically, the proposed method extends the two-way wave-equation-based wavefield extrapolation framework from acoustic media to anisotropic VTI media. To further improve image quality, we used a frequency–wavenumber-domain approach to remove pseudo-S-wave energy, enhancing the clarity of P-wave descriptions and reducing imaging artifacts. We evaluated the effectiveness of the proposed approach through several numerical tests, including experiments on a VTI three-layer model, the Hess VTI model, and the Marmousi VTI model. These benchmarks demonstrated that our method consistently outperformed conventional acoustic schemes designed for isotropic media, yielding clearer, higher-resolution structural images. In addition, we applied both the acoustic and the VTI-based migration methods to a real seismic dataset. The VTI-based migration provided a clearer, more continuous image of the deep structures than the acoustic migration method. The numerical experiments and real data application demonstrated that the proposed VTI imaging strategy is both practical and effective. More importantly, the findings underscore the need to explicitly account for anisotropic parameters in seismic imaging workflows to achieve geologically consistent and reliable interpretations in VTI-dominated regions.

Surface waves are a prevalent form of coherent interference in seismic recordings, characterized by low frequency, high energy, and slow velocity, which can significantly affect seismic data processing. Currently, surface wave suppression is primarily achieved by leveraging the differences between surface waves and signal waves in the frequency, amplitude, and wavenumber domains. Among these methods, frequency–wavenumber (FK) filtering is widely used. However, it requires manual selection of regions during filtering, which becomes challenging when processing large volumes of seismic data. Therefore, adaptive surface-wave suppression methods are essential for practical data processing. FK filtering transforms data into the FK domain to suppress linear noise. However, because surface waves are dispersive and exhibit partial nonlinearity, FK filtering alone is insufficient to fully suppress them. To address these issues, this study introduces a temporal dimension into the FK domain, leading to the development of the time–frequency–wavenumber (TFK) transformation. This transformation further separates surface and signal waves in the time domain, thereby mitigating dispersion. Moreover, it facilitates adaptive filtering of seismic data across different time intervals in the FK domain. By cross-correlating FK data from different time periods, adaptive filters were derived for each interval, which were then applied to obtain filtered seismic records. Comparisons between synthetic and real-world data demonstrated that our approach effectively suppresses surface waves while preserving the relative amplitude characteristics of signal waves.

The mudstone in a specific exploration area of the Yinggehai Basin is rich in organic matter and serves as a paradigmatic example of a low-velocity mudstone formation in rock physics. This type of interval exhibits seismic response characteristics similar to those of hydrocarbon reservoirs, which complicates the identification of gas reservoirs and the prediction of gas-bearing zones in this area. In this study, the analysis and modeling of rock physical characteristics within the exploration area are investigated. Based on the actual drilling curve and coring analysis data, a rock physics model of the low-velocity mudstone reservoir is established using self-consistent approximations and the Ciz–Gassmann model, and the influence of organic matter content on the mudstone’s elastic properties is analyzed. Additionally, a robust quantitative inversion procedure is introduced to test the feasibility of inverting for porosity, clay content, and water saturation to mitigate the risk of low-velocity mudstone. The application of actual data, including core samples, logging data, and seismic database, demonstrates the effectiveness of this method and provides technical support for seismic prediction of sand body identification and hydrocarbon detection.

Conventional ant-tracking technology is effective for fault identification, but its application is constrained by sensitivity to data quality and inadequate continuity of fault tracking results. This limitation creates an urgent need for optimization. To address this issue, this study proposes a frequency-divided ant-tracking method based on W-transform inversion spectral decomposition (WT-ISD) to improve the accuracy of fracture prediction in complex structural areas. By integrating the flexible time–frequency localization of the W-transform with sparse inversion theory, this method markedly improves the resolution and focusing performance of time–frequency spectra. This is achieved through the construction of an overcomplete dictionary and the imposition of L1 norm constraints. The proposed method not only provides a novel high-resolution time–frequency analysis tool for seismic signal processing but also establishes a theoretical basis for subsequent fine geological interpretation via its frequency-divided processing workflow. Taking the Raphia S Block in the Bongor Basin as a case study, this research first adopts WT-ISD to obtain high-resolution spectral decomposition results, generating single-frequency data volumes corresponding to different frequency components. Subsequently, structural smoothing filtering and boundary enhancement processing are applied to highlight discontinuity information. Then, the ant-tracking algorithm is introduced, combined with regional structural attitude constraints, to realize the identification of multi-scale fractures. Finally, red–green–blue attribute fusion technology is used to integrate responses from different frequencies and construct a fault distribution model. Practical application results indicate that this method not only improves the accuracy and spatial continuity of fracture prediction but also enhances its anti-noise capability. In particular, it exhibits excellent identification performance for large-, medium-, and small-scale fractures corresponding to 20 Hz, 35 Hz, and 50 Hz, respectively. This study verifies that the proposed method can provide reliable technical support for multi-scale fracture detection in complex structural areas and demonstrates important application value for fracture prediction in oil and gas exploration.

Irregular topography can generate out-of-plane signals (OPS) on seismic sections, interfering with the imaging of the true seafloor directly beneath the survey line. While acquiring three-dimensional data or using specialized sensors can mitigate this, these options are often costly or unavailable, especially for legacy surveys. To efficiently remove OPS from two-dimensional (2D) data, this study investigates the validity of using a neural network (NN) for picking and muting. First, we demonstrate the limitation of conventional frequency–wavenumber domain directional filtering due to the kinematic similarity between OPS and true seafloor reflections. Then, we present a workflow that employs a cascade–correlation learning algorithm to identify and mute OPS arrivals before the first break. Unlike data-intensive deep learning techniques that require large training datasets, this lightweight NN is trained on user-picked examples of true seafloor reflections, enabling it to distinguish OPS events arriving from outside the vertical survey plane. Application of this technique to a 2D line acquired near irregular seafloor topography in the Ulleung Basin demonstrates the true seafloor reflector and the removal of false offline signals. Qualitative and quantitative validation against an independent external bathymetric reference both showed a reduction in travel time error compared to the raw data, confirming the effectiveness of the picking results. The results highlight that a cascade–correlation NN-based picking and muting can efficiently suppress OPS in cases of irregular topography on 2D seismic data.