A Seismic Foundation Model for Dense Feature Extraction
Authors: Zhixiang Guo, Xinming Wu, Yimin Dou, Hui Gao, Siyao Zhao, He Huang, Wei Liu
Organizations: Laboratory of Seismology and Physics of the Earth’s Interior, School of Earth and Space Sciences, University of Science and Technology of China, Hefei 230026, China. · State Key Laboratory of Precision Geodesy, University of Science and Technology of China, Hefei 230026, China. · Mengcheng National Geophysical Observatory, University of Science and Technology of China, Hefei 230026, China. · School of Artificial Intelligence, Anhui University of Science and Technology, Hefei 231131, China
Deep-learning-based seismic interpretation is constrained by scarce high-quality annotations, and most existing models are trained separately for individual tasks. We present SeisDFE, a 7B-parameter seismic foundation model for dense feature extraction, pretrained on 5.34 million 2D images derived from more than 500 seismic volumes. A three-stage pretraining strategy, combining discriminative self-supervised learning, Gram feature anchoring, and multi-resolution adaptation, improves feature discriminability while preserving spatial coherence. With its encoder frozen, SeisDFE supports seismic facies classification, salt detection, fault detection, and relative geologic time estimation. It outperforms existing seismic foundation models on all four tasks and a natural-image-pretrained model on all but fault detection, with the largest gains under scarce supervision: using only two annotated sections of the F3 dataset, it improves whole-volume mIoU by 12-19 percentage points over the baselines. The learned features remain effective with task-specific decoders and extend to synthetic-to-field transfer and sparse 3D geobody interpretation. These results show that large-scale seismic pretraining organizes geological information into dense representations that are directly reusable across interpretation tasks.
Figures & tables
Figure 1: Overall framework of SeisDFE. The left panel shows the three-stage pretraining pipeline, consisting of large-scale self-supervised learning, Gram feature anchoring, and multi-resolution adaptation. The middle panel shows the frozen ViT-7B seismic foundation model and its patch-level dense features. The right panel shows the adaptation of task-specific decoders for seismic facies classification, fault detection, salt detection, RGT estimation, and other interpretation tasks.
Figure 2: The discriminative self-supervised pretraining framework of SeisDFE. Each seismic image yields two global views and eight local views, which are fed to the teacher and student encoders respectively, with the teacher parameters updated by an exponential moving average of the student parameters. The training objective comprises the DINO loss, the iBOT loss, the KoLeo regularizer, and the Gram anchoring loss.
Figure 3: Evolution and organization of dense seismic representations. (a) Patch-feature similarity during Stage I at 50k, 80k, and 100k training steps. (b) PCA visualization of dense features after the three pretraining stages. (c) Frozen linear-probing performance of the three stages on SEAM facies classification. (d) PCA visualization of features extracted from different depths of the final SeisDFE encoder.
Figure 4: Extreme few-shot linear probing with frozen encoders. (a) Salt detection, (b) F3 facies classification, (c) RGT estimation, and (d) fault detection. Only a linear prediction head is trained for each task.
Task
Decoder
Encoder
Encoder setting
mIoU
mPA
SEAM facies classification
PUP
DINOv3-7B
Frozen
0.8374
0.9221
SFM-Large
LoRA
0.8243
0.8994
NCS-2D
Frozen
0.8415
0.9121
SeisDFE-Large
Frozen
0.8768
0.9289
SeisDFE
Frozen
0.8845
0.9324
TGS salt detection
Linear
DINOv3-7B
Frozen
0.8116
0.8815
Table 1: Performance of different pretrained encoders on three downstream seismic interpretation tasks. Bold indicates the best result.
Figure 5: Qualitative comparisons for downstream seismic interpretation. The rows show seismic facies classification, salt detection, and fault detection. The columns show the seismic input, predictions from DINOv3-7B, SFM-Large, NCS-2D, and SeisDFE, and the reference interpretation.
Figure 6: Synthetic-to-field transfer on field seismic data. The upper two rows show two examples each of fault detection (left) and karst cave detection (right), with predictions obtained using SeisDFE and U-Net in the first and second rows, respectively. The bottom row shows two RGT estimation examples, each comparing SeisDFE (left) and U-Net (right).
Figure 7: Sparse 3D geobody detection. (a) Overview of the 3D interpretation framework, in which multi-level features from neighboring seismic sections are combined and processed by a 3D convolutional decoder under sparse slice supervision. (b) Sparse manual interpretations used for supervision. (c) Predictions obtained with a 2D PUP decoder. (d) Predictions obtained with the 3D convolutional decoder. The four columns in (b)–(d), from left to right, show channel detection on the Parihaka and Romney datasets, salt dome detection on the F3 dataset, and karst detection on a field seismic dataset.
Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.
Laura Laurenti, David Liu, Brian Kulis +2
Institute of Geophysics, ETH Zurich · Swiss Seismological Service, ETH Zurich · Dept. of Computer Science, Boston University +2
The advent of large pretrained foundation models for computer vision has significantly improved the efficiency of visual data interpretation. The Segment Anything Model (SAM), in particular, offers powerful zero shot segmentation capabilities through prompt based interaction, thus making it a promising tool for seismic interpretation. However, most existing applications of SAM rely on fine tuning for specific geological targets, which requires extensive labeled data, incurs high computational cost, and often compromises the model's generalization capability. In this study, we introduce a principled framework for zero shot adaptation of foundation models to seismic data. The framework is built on two key components: (1) aligning seismic attributes and visualization choices (e.g., colormaps) with the geological target of interest, and (2) employing a hybrid prompting strategy that combines sparse user defined point prompts with dense mask prompts derived from SAM's internal feature activations. We systematically evaluate this framework across multiple geological targets, datasets, prompt configurations, and seismic attribute representations. Our results demonstrate that geologic target aware selection of seismic attributes and colormaps, combined with hybrid prompting, enhances the separability of geological features and improves boundary delineation and segmentation accuracy relative to point based prompting alone. Our findings show that, when these components are jointly applied, SAM can achieve competitive segmentation performance in a fully zero shot setting, thereby eliminating the need to retrain SAM for each geologic feature. This work establishes a practical and scalable pathway to leverage foundation models in seismic interpretation, reducing reliance on labeled data while preserving model generality.
Aniq Ahmad, Heather Bedle, Ahmad Mustafa
School of Geosciences, University of Oklahoma, Norman, Oklahoma, USA · King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
The demand for high-resolution subsurface imaging and continuous Earth monitoring has driven rapid growth in active and passive seismic data from dense geophone deployments, distributed acoustic sensing (DAS) arrays, and large-scale 2D and 3D surveys. This expansion makes complex noise suppression increasingly challenging, especially when signal fidelity must be preserved. Conventional supervised deep learning methods are often task-specific, require large paired datasets, and can suffer from domain shift under new acquisition conditions. Foundation models offer a promising alternative, but pre-training seismic foundation models from scratch requires massive domain-specific data and substantial computation. We propose an efficient framework that repurposes general-purpose Vision Foundation Models (VFMs) for geophysical tasks through Parameter-Efficient Fine-Tuning. The architecture uses a pre-trained VFM, a DINOv3 encoder, adapted with Low-Rank Adaptation (LoRA) to enable effective feature adaptation with few additional parameters. To improve robustness under unseen field conditions without ground truth, we introduce a kurtosis-guided unsupervised test-time adaptation module that updates only LoRA parameters during inference. This module self-calibrates the model to site-specific noise by identifying information-rich regions via kurtosis and performing self-training without labeled data. Experiments on public exploration seismic images and DAS vertical seismic profiling data from the Utah FORGE site show that the framework matches or outperforms domain-specific models. Tests on unseen cross-site data from a land survey in China and the Groß Schönebeck geothermal site in Germany further demonstrate strong generalization and effective signal-noise separation. These results highlight the potential of adapting pre-trained VFMs to data-intensive problems in exploration seismology.
Jiahua Zhao, Umair bin Waheed, Jing Sun +3
Computation-based Science and Technology Research Center, The Cyprus Institute, Nicosia, Cyprus · Department of Geosciences, College of Petroleum Engineering and Geosciences, King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia · Department of Intelligent Systems, Faculty of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology, Delft, Netherlands +2