Abstract Seismic data obtained after stacking or migration typically exhibit low dominant frequency, sparse spatial and temporal sampling, and may retain a low signal-to-noise ratio (SNR) even after stacking. To enhance the resolution of seismic data, we propose a deep learning-based single-step super-resolution (SR) method that simultaneously performs denoising, interpolation, and frequency extrapolation. We design a seismic data super-resolution Transformer (SDST) to extract and fuse local and global features from low-resolution seismic data, with particular emphasis on the amplitude characteristics and spatial continuity of effective events (e.g., horizons, folds, faults). Edge information from the input data is extracted using the Canny algorithm and used as a mask to generate an edge map, which is then concatenated with the raw data to form a two-channel input, thereby improving amplitude preservation and noise suppression. The SDST backbone comprises residual groups containing Swin Transformer layers which capture features within windows while reducing the computational cost of self-attention by restricting attention to finite-sized windows. Pixel shuffle layers are employed to efficiently upsample the output feature maps from the SDST backbone, thereby enhancing the details of geological structures. In addition, upsampling the input data through a skip connection helps preserve the output amplitudes, particularly for weak events. We prepare 3-dimensional (3D) synthetic seismic volumes with complex geological structures, half of which contain a mixture of strong- and weak-amplitude events. 2-dimensional (2D) slices are then randomly selected from these volumes to construct training datasets that better approximate the characteristics of field data for supervised learning. Numerical tests on both synthetic and field data from different exploration regions demonstrate the feasibility of the proposed method.
Abstract Due to the growth of data volumes in monitoring seismicity, automation is crucial to fully use the information content. Seismic event detection is the first step in processing, aimed at reducing the amount of data by extracting intervals with seismic activity. At the same time, interpretability remains important for justification and quality control of the detected events, especially when multiple candidate events are to be triaged. A novel seismic event detection technique based on cluster analysis of trimmed spectrograms (CATS) is presented. It uses an automatic noise estimator, B-E-DATE, and cluster filtering in the time-frequency domain. CATS workflow is fully transparent, offering a framework for comprehensive quality control. Evaluated on the benchmark dataset TXED, CATS showed competitive performance compared to traditional and deep learning methods by showing high recall, precision, and F-scores. Furthermore, CATS can be extended to multi-station processing and employed directly for denoising without requiring any extra parameters, and its outputs can be utilized for phase-picking and event classification.
Abstract Efficient and accurate magnetotelluric (MT) forward modeling is essential for three-dimensional MT inversion and data interpretation. While existing MT forward solvers on unstructured tetrahedral meshes predominantly employ direct solvers, the development of efficient iterative methods for high-order elements remains significantly underexplored compared to first-order elements, despite their superior accuracy. To address this, we present a unified iterative framework for both first- and second-order finite element (FE) formulations. The complex-valued FE system is transformed into an equivalent real-valued form and solved using the flexible generalized minimal residual (Flexible GMRES, FGMRES) method with an optimal block-diagonal preconditioner. This preconditioning is performed by the auxiliary-space Maxwell solver (AMS) based on the Hiptmair–Xu decomposition. To eliminate spurious divergence and accelerate convergence, particularly for models with complex topography, an explicit divergence projection (EDP) scheme is introduced to filter out non-physical components from the intermediate preconditioned vectors during FGMRES iterations. This is achieved by solving a Poisson equation for a scalar potential using the preconditioned conjugate gradient method with an algebraic multigrid preconditioner, utilizing the divergence of the conductivity-scaled vector as the source term. Numerical experiments validate our algorithm against established open-source finite difference and FE solvers, confirming that both FE approaches achieve the required accuracy, with the second-order scheme generally yielding smoother and more precise solutions. Results indicate that at lower frequencies, AMS-based preconditioning becomes sensitive to air conductivity, where excessively low values can trigger numerical instabilities that compromise solver convergence. Consequently, an appropriate choice of air conductivity is critical for sustaining numerical stability and superior solver performance. Furthermore, while complex topography typically degrades solver efficiency, the EDP technique mitigates this performance penalty by suppressing the accumulation of spurious divergence. This stabilization of iteration counts ensures robust and efficient convergence across a broad frequency range, even in the presence of challenging terrain.
Abstract The objective of this study is to understand the effect of height on ground penetrating radar (GPR) data for the near surface ground-coupled regime. The height range investigated was from 0.017 to ~0.083 of the GPR's center frequency wavelength in free space (λc). A shielded 250 MHz center frequency GPR placed above a homogeneous medium, with a perfect electrical conductor (PEC) layer located at its base, was numerically modeled using a three-dimensional (3D) ?nite-difference time-domain (FDTD) algorithm to understand the GPR's response dependence on height for relative dielectric permittivity (Kr) values from 4.5 to 18. The numerical modeling results show that the optimal height to minimize direct wavelet and banding interference is ~0.033- 0.050 λc (i.e., 40-60 mm at 250 MHz). Elevating the GPR above the surface also increases airwave interference from above surface objects. The numerical modeling results show the dependence of this interference on GPR height and the orientation of the above surface planar objects relative to the GPR. Field site profiles were collected over a subsurface utility with a commercial 250 MHz center frequency GPR at a range of heights. These profiles show some similar height dependence characteristics and some differences with the numerical modeling results. Since shields, carts and antenna designs vary with GPR manufacturers, data quality height dependence should be field tested to determine an optimal height. The numerical model and field data results provide an understanding of the tradeoffs in choosing an optimal height for GPR profiling, based on survey objectives.
Abstract Deep learning-based 3D inversion of gravity and magnetic field data can directly predict the spatial distribution and physical property parameters of subsurface source bodies from observed data, offering high resolution and efficiency. However, in practical applications, its generalization capability is constrained by two factors: the extent to which the training dataset approximates real geological conditions, and the network's capacity to extract data features. Previous model construction, forward modeling, and inversion methods are reviewed and compared. First, to tackle the problem that existing modeling methods cannot simulate rock bodies with gradational morphologies and physical properties, the continuous smoothness characteristics of Perlin noise are introduced for 3D geophysical model construction, enabling the simulation of realistic geometric shapes and physical properties of source bodies. Second, to address the dataset construction efficiency issue, a fast surrogate network is proposed, significantly accelerating batch forward modeling while maintaining precision. Third, a novel network integrating convolutional neural networks and Transformers is designed, combined with a dynamic noise-adding strategy, effectively enhancing the resolution of inversion results. In addition, to better quantify and evaluate inversion reliability, a new evaluation metric is proposed that characterizes both spatial and physical property consistency between the predicted model and the true model. Subsequently, a deep learning inversion framework is proposed by integrating the above methods, aiming to enhance the generalization ability and practicality of deep learning-based 3D inversion of gravity and magnetic data. Tests on synthetic data demonstrate the effectiveness of the proposed methods. Finally, the generalization capability of the inversion method and the effectiveness of the evaluation metric are discussed, and the method is validated on real data from the San Nicolas deposit in Mexico, yielding a subsurface 3D density model that is consistent with available geological and borehole information, and confirming the effectiveness and feasibility of the method.
