In oil and gas seismic exploration, fluid factors are key parameters for identifying reservoir fluid properties and evaluating reservoir potential. Although regularization methods are commonly used to enhance inversion stability, traditional time-domain lateral constraint methods struggle to effectively address the issue of abrupt lateral stratigraphic variations. This paper aims to improve the prediction accuracy and stability of fluid factors in such scenarios. Based on the characteristic that frequency-domain seismic data exhibit stronger correlation during stratigraphic abrupt changes, this paper proposes a frequency-domain two-step sub-band regularization inversion method. First, a difference operator is introduced into the frequency domain to construct the objective function, and the Alternating Direction Method of Multipliers algorithm is adopted as the solution. Furthermore, a two-step sub-band regularization strategy is proposed: First, integrating low-frequency and high-frequency residual terms into the same objective function; then performing inversion in two stages. These stages first use low-frequency data inversion to obtain the general profile of the subsurface structure as an initial model, followed by using high-frequency data inversion based on this to obtain detailed information. Theoretical model tests have verified the superiority of this method under complex geological conditions. Field application in practical work areas shows that the frequency-domain two-step method significantly outperforms the traditional time-domain lateral constraint L1 regularization method in terms of vertical resolution and lateral continuity. This method provides a more accurate and stable solution for fluid identification under complex geological conditions.

As a World Cultural Heritage Site, the Yungang Grottoes face the risk of geological disasters caused by underground goafs. To accurately detect the surrounding void zones of the grottoes and analyze their distribution and structural characteristics, this study uniquely applied the passive-source surface wave spatial autocorrelation (SPAC) method. Four high-density linear arrays were deployed in the Yungang Grottoes area to collect microtremor signals. By extracting the Rayleigh wave dispersion curve and combining it with a non-linear least-squares inversion technique incorporating a damping regularization term, a shallow shear wave velocity profile was established, successfully identifying potential low-velocity anomalies. During the inversion process, the optimized damping factor effectively suppressed the oscillation effects under complex geological conditions, significantly improving the stability and accuracy of the imaging. The results showed that 15 typical low-velocity anomalies with wave velocities below 1,200 m/s were identified, and these anomalies were highly consistent with the locations of goaf areas in historical mining data. The imaging results revealed that the goaf exhibits multiple layers and a multi-center distribution, with the L1 and L4 survey lines being the main areas of goaf activity. With a horizontal resolution of 10-20 m and a maximum detection depth of 320 m, the method presented differences in the integrity and fragmentation of underground rock masses. The SPAC method demonstrated advantages such as high resolution, non-destructive testing capability, and imaging stability in the detection of abandoned mine areas within cultural heritage sites. By optimizing the inversion regularization parameters, this study significantly improves imaging accuracy in complex geological environments, providing effective technical support for the stability assessment of the Yungang Grottoes and for the prevention and control of geological disaster risks at cultural heritage sites.

The resolution of seismic images significantly impacts the accuracy of subsequent seismic interpretation and reservoir location. However, the resolution of seismic images often degrades due to the influence of multiple factors, making super-resolution of seismic images essential and critical. We propose a grouped-residual and multi-scale large-kernel attention network (GMLAN) framework, trained on synthetic seismic images to achieve excellent seismic image super-resolution on field seismic data. GMLAN is primarily composed of two modules: The feature extraction module (FEM) and the image reconstruction module (IRM). The FEM consists of two components: Shallow feature extraction (SFE) and deep feature extraction (DFE). The SFE component is designed to capture the basic information of seismic images, such as large-scale structures and morphological features of the strata. The DFE component serves as the cornerstone of the feature extraction process, leveraging residual groups and multi-scale large-kernel attention to distill detailed features from seismic images, such as stratigraphic interfaces, dip angles, and relative amplitudes. Finally, the IRM utilizes sub-pixel convolution, a learnable upsampling technique, to reconstruct super-resolution seismic images while preserving the continuity of seismic features. The framework demonstrates satisfactory performance on both synthetic and field data.

Distributed acoustic sensing (DAS) has attracted much attention in seismic data acquisition because of its low cost, anti-electromagnetic interference, and high acquisition density. Unfortunately, the acquired DAS records are usually accompanied by various kinds of complex noise, affecting subsequent interpretation and inversion. Traditional methods have difficulties in effectively attenuating the intense background noise. In general, the denoising task of DAS data is challenging. Recently, convolutional neural networks (CNNs) exhibit a good ability in suppressing the noise in DAS records. However, traditional CNN-based frameworks always have a relatively simple network architecture, bringing negative impacts on the denoising capability. To solve this problem, we propose a dual-branch dense network (DBD-Net) in this paper. Specifically, DBD-Net introduces a novel combination of dual-branch modules and an attention mechanism: the dual-branch modules extract multi-scale coarse-to-fine features, while the attention mechanism highlights the most informative features. This joint design strengthens feature representation and signal recovery compared with conventional CNN structures such as denoising CNN (DnCNN) and U-Net. Moreover, an attention module is employed to enhance the effective features. To verify the denoising ability, we compare DBD-Net with other competing methods, including band-pass filter, DnCNN, and U-Net, in terms of denoising capability and processing accuracy. Experimental results verify that DBD-Net can improve the quality of DAS records with a signal-to-noise ratio increment of nearly 26 dB. Meanwhile, the intense DAS background noise is also perfectly suppressed and the weak signals are effectively restored, representing advantages over the competing methods.

Towed-streamer marine seismic acquisition systems generally have a dense receiver spacing in the inline receiver direction within common-shot gathers (along-streamer), while the streamer spacing is relatively sparse in the crossline receiver direction within common-shot gathers (cross-streamer). This disparity can lead to spatial aliasing issues in the crossline receiver direction within common-shot gathers and result in resolution degradation during the processing of 3D seismic data. To address this issue and enhance resolution, data interpolation in the crossline receiver direction within common-shot gathers is essential. Various supervised learning-based interpolation methods have been developed to this end. However, the absence of true data in the crossline receiver direction within common-shot gathers poses challenges for training supervised learning models with actual field data. To overcome this, we have developed a novel approach called the “transposed arrangement strategy” for a deep learning-based reconstruction model for crossline interpolation. This method involves training the model with 3D input and labels patched from existing field data, and then applying the trained model with transposed 3D input to reconstruct data in the crossline receiver direction within common-shot gathers. During this process, the 3D U-Net and U-Net+ models were utilized, demonstrating their superiority through comparisons with traditional interpolation methods.

Seismic impedance inversion is essential for reservoir characterization but remains challenging in complex geological environments due to the inherent limitations of conventional methods. This study proposes a hybrid deep learning framework integrating a convolutional neural network (CNN), a graph attention network (GAT), and a gradient boosting decision tree (GBDT) to achieve high-resolution impedance inversion. The CNN extracts local structural features from seismic waveforms, the GAT captures long-range geological dependencies through self-attention between traces, and the GBDT performs robust non-linear regression for final prediction. Extensive evaluations on synthetic and field datasets demonstrate that the method achieves a root mean square error of 285 m/s·g/cm3 on the Society of Exploration Geophysicists salt model, representing a 15.2% improvement over XGBoost and a 32.1% improvement over sparse spike inversion. The framework performs particularly well in complex regions, achieving a 22.7% error reduction at salt boundaries and a thin-bed detection rate of 92% for layers exceeding 4 m in thickness. Statistical uncertainty quantification indicates 94.2% coverage of true impedance values within 95% confidence intervals. In practical applications, the method reduces interpretation time by 40% while maintaining reservoir thickness prediction errors within ± 3 m, demonstrating strong robustness and operational value for seismic interpretation.