ORIGINAL RESEARCH article
Front. Earth Sci.
Sec. Geoinformatics
Leakage-Aware Full-Section Evaluation and Label-Efficient Semi-Supervised Seismic Facies Segmentation for Subsurface Geological Modeling
- ZL
Zhibin Li 1,2,3
- JL
Jingxian Liu 1,2,3,4
- XL
Xuece Li 1,2,3
- TH
Tongtong He 1,2,3,4
- KL
Kangning Li 1,2,3
- YZ
Yuanning Zhang 1,2,3
- ZP
Zhongkui Pan 1,2,3
- XL
Xiangjun Lei 1,2,3
- SW
Shuheng Wang 1,2,3
- ZZ
Zhaojia Zhang 1,2,3
1. The Third Institute of Geology and Mineral Exploration, Gansu Provincial Bureau of Geology and Mineral Exploration and Development, Lanzhou, China
2. Gansu Provincial Technology Innovation Center for Gold Resource Exploration and Utilization, Lanzhou, China
3. Gansu Provincial Key Laboratory of Mineral Resource Exploration, Lanzhou, China
4. Lanzhou University School of Earth Sciences, Lanzhou, China
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
Abstract
Seismic facies segmentation assigns facies labels to 3D seismic data and supports subsurface geological modeling. Yet reported performance is often weakened by spatial leakage from random patch splits, single-survey validation, and inconsistent training recipes. We build a leakage-aware full-section evaluation framework. It treats each inline or crossline section as one sample. On Netherlands F3, it uses spatially contiguous training/validation splits with buffer gaps and evaluates on independent test cubes. Under one training recipe and three seeds, we benchmark 11 CNN and Transformer architectures on F3. We also study spatial nonstationarity and matched-support low-label learning on New Zealand Parihaka, where no official independent test cube is available. The results show that architecture choice alone is not the main performance source. Pretraining, backbone capacity, and survey nonstationarity matter more. Based on this finding, we propose a label-efficient semi-supervised method without changing the network architecture. It uses weak-to-strong consistency on unlabeled sections and seismic-domain strong augmentation for amplitude, noise, and spectral changes. Logit-adjusted pseudo-labeling is added as an optional rare-class component. Experiments show that the basic weak-to-strong method outperforms paired supervised baselines on both surveys at 5%, 10%, and 25% label ratios. With 5% labels, mIoU improves by 0.061 on F3 and by 0.181 on Parihaka. Logit adjustment improves rare classes on F3, but can over-correct on nonstationary Parihaka. The study provides an independent-cube F3 benchmark and a robust low-label training practice for seismic facies segmentation.
Summary
Keywords
full-section protocol, leakage-aware evaluation, Pseudo-labeling, seismic facies segmentation, Semi-Supervised Learning, Spatial nonstationarity, Subsurface geological modeling, weak-to-strong consistency
Received
16 June 2026
Accepted
13 August 2026
Copyright
© 2026 Li, Liu, Li, He, Li, Zhang, Pan, Lei, Wang and Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Jingxian Liu
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.