3D seismic salt dome segmentation using a hybrid UNet with global attention, denoising pretraining, and discriminative transfer learning
- 1Corresponding author.
Abstract
Automated delineation of salt structures from three-dimensional seismic data remains a fundamental challenge in structural interpretation, constrained by severe scarcity of expert-labeled training data and the persistent tendency of segmentation networks to misclassify acquisition noise as salt boundaries. We present the first application of unfactorized three-dimensional global self-attention to seismic salt dome segmentation, within a hybrid UNet combining volumetric convolution with complete spatial context modeling. The model is trained with self-supervised denoising pretext training across three unlabeled surveys, followed by discriminative transfer learning on binary Zechstein salt annotations from the Netherlands F3 block using layer-wise learning rate decay, gradual unfreezing, and Unified Focal Loss. This pretext task develops explicit noise-discrimination representations before salt-specific supervised adaptation; in a controlled comparison under the identical architecture, it yielded a larger gain over random initialization than contrastive (Barlow Twins-style) or masked-reconstruction pretext tasks. Validation on the F3 dataset achieved Dice = 0.843, IoU = 0.729, Precision = 0.831, and MCC = 0.829, representing statistically significant improvements of 5.0–9.2% over the strongest comparison model, the 3D CNN + Swin architecture, and exceeding a Salt3DNet-style self-supervised reference model on all six metrics. On the spatially distinct blind test partition containing a large Zechstein diapir, performance gaps over Swin widened to 10.6–18.1%, with the proposed model exhibiting substantially lower degradation (3.7% Dice reduction) versus Swin (8.6%) and the baseline CNN (11.7%). The proposed model remains compact and memory-efficient, with 3.43M parameters (3.18M for the baseline) and only 0.9 GB of GPU memory for training thanks to fused attention, while exact global attention over the full 64 3 sub-volume requires 6.8 TFLOPs per forward pass. Zero-shot cross-basin evaluation on the recently released Gulf of America OCS volume (NAMSS survey B-51-94-LA), assessed qualitatively on a survey with no prior salt interpretation, provides the first automated structural delineation of Louann Salt diapir geometries in this survey; the results are qualitatively consistent with generalization to geologically distinct salt morphologies and constitute an original geoscientific contribution to one of the world's most tectonically significant salt provinces.
Related
- Direction-Aware Masked Pretraining on 3D Seismic Data with Transfer to Cross-Area Acoustic Impedance Inversion
- Fault-conditioned Seismic Image Generation using Denoising Diffusion Probabilistic Modeling and Neural Style Transfer
- HyEff-UNet: an attention-driven Efficientnet U-Net with hybrid loss optimization for UAV-based marine oil spill detection and segmentation
- Plume Segmentation from MethaneSAT with Cross-Sensor Transfer Learning and Physics-Informed Postprocessing
- Xaurora: Generative Weather Forecasting with Denoising Stochastic Interpolants from a Foundation Model Prior
BibTeX
@article{mahzad2026seismic,
title = {{3D seismic salt dome segmentation using a hybrid UNet with global attention, denoising pretraining, and discriminative transfer learning}},
author = {Matin Mahzad and Majid Bagheri},
journal = {Applied Computing and Geosciences},
year = {2026},
month = nov,
doi = {10.1016/j.acags.2026.100424},
url = {https://doi.org/10.1016/j.acags.2026.100424},
note = {CC BY},
}
RIS
TY - JOUR
TI - 3D seismic salt dome segmentation using a hybrid UNet with global attention, denoising pretraining, and discriminative transfer learning
AU - Matin Mahzad
AU - Majid Bagheri
PY - 2026
DA - 2026/11/08
JO - Applied Computing and Geosciences
DO - 10.1016/j.acags.2026.100424
UR - https://doi.org/10.1016/j.acags.2026.100424
AB - Automated delineation of salt structures from three-dimensional seismic data remains a fundamental challenge in structural interpretation, constrained by severe scarcity of expert-labeled training data and the persistent tendency of segmentation networks to misclassify acquisition noise as salt boundaries. We present the first application of unfactorized three-dimensional global self-attention to seismic salt dome segmentation, within a hybrid UNet combining volumetric convolution with complete spatial context modeling. The model is trained with self-supervised denoising pretext training across three unlabeled surveys, followed by discriminative transfer learning on binary Zechstein salt annotations from the Netherlands F3 block using layer-wise learning rate decay, gradual unfreezing, and Unified Focal Loss. This pretext task develops explicit noise-discrimination representations before salt-specific supervised adaptation; in a controlled comparison under the identical architecture, it yielded a larger gain over random initialization than contrastive (Barlow Twins-style) or masked-reconstruction pretext tasks. Validation on the F3 dataset achieved Dice = 0.843, IoU = 0.729, Precision = 0.831, and MCC = 0.829, representing statistically significant improvements of 5.0–9.2% over the strongest comparison model, the 3D CNN + Swin architecture, and exceeding a Salt3DNet-style self-supervised reference model on all six metrics. On the spatially distinct blind test partition containing a large Zechstein diapir, performance gaps over Swin widened to 10.6–18.1%, with the proposed model exhibiting substantially lower degradation (3.7% Dice reduction) versus Swin (8.6%) and the baseline CNN (11.7%). The proposed model remains compact and memory-efficient, with 3.43M parameters (3.18M for the baseline) and only 0.9 GB of GPU memory for training thanks to fused attention, while exact global attention over the full 64 3 sub-volume requires 6.8 TFLOPs per forward pass. Zero-shot cross-basin evaluation on the recently released Gulf of America OCS volume (NAMSS survey B-51-94-LA), assessed qualitatively on a survey with no prior salt interpretation, provides the first automated structural delineation of Louann Salt diapir geometries in this survey; the results are qualitatively consistent with generalization to geologically distinct salt morphologies and constitute an original geoscientific contribution to one of the world's most tectonically significant salt provinces.
SN - 2590-1974
ER -