Skip to main content
Back to timeline
arXivSource publication:

SeisDFE, a 7B seismic foundation model, works across four tasks with a frozen encoder: two annotated F3 sections reach 0.82 whole-volume mIoU

Synopsis

Seismic interpretation is constrained by scarce high-quality annotations and by models trained separately for each task. SeisDFE, a 7B-parameter model pretrained on 5.34 million 2D images from more than 500 seismic volumes, uses three-stage pretraining with discriminative self-supervision, Gram feature anchoring and multi-resolution adaptation, and with its encoder frozen supports facies classification, salt detection, fault detection and relative geologic time estimation. Using only two annotated F3 sections, it reaches a whole-volume mIoU of 0.82, 12 to 19 percentage points above the baselines, with the largest gains under the scarcest supervision.

Source-provided article image: A Seismic Foundation Model for Dense Feature Extraction
arXiv

Interpretation

A single frozen seismic encoder can now serve four interpretation tasks instead of relearning features for each target. Most prior seismic deep-learning workflows learned representations separately for each target and dataset, so the same structural information was relearned from limited annotations. On F3 facies data, training a linear head on only two annotated sections and inferring over the full 3D volume, SeisDFE reaches a whole-volume mIoU of 0.82, versus 0.70 for DINOv3-7B, 0.66 for NCS-2D and 0.63 for SFM-Large.

Pretraining proceeds in three stages that turn visual features into seismic dense representations that are both discriminative and spatially coherent. Discriminative self-supervised training degrades local spatial relations late in optimization, and reconstruction-based objectives do not explicitly require geological patterns to be separable or patch relations stable. Stage I trains 100k steps on 5.34 million images, Stage II adds 10k steps with the 50k-step Stage I checkpoint as a frozen Gram teacher constraining patch-feature cosine-similarity matrices, and Stage III adds 30k steps of multi-resolution adaptation on the retained 600,000 images; under the same frozen linear probe, SEAM facies mIoU rises from 0.6332 after Stage I to 0.6431 after Stage II and 0.6530 after Stage III.

Perspective

Evaluation covers four 2D tasks: salt detection, F3 facies classification, fault detection and relative geologic time estimation, with the encoder frozen in both few-shot linear probing and task-specific decoder settings; synthetic-to-field transfer and sparse 3D geobody interpretation are compared only qualitatively.

The field fault, karst and relative-geologic-time experiments lack complete reference labels and provide only qualitative evidence of synthetic-to-field transfer; the 3D interpretation experiments use only sparsely annotated slices, limiting quantitative assessment of volumetric accuracy; the compared foundation models differ in architecture, parameter scale, pretraining data and adaptation strategy, which complicates attributing gains to individual pretraining components.