Deep learning driven horizon tracking with confidence
File(s)
Published version
Author(s)
Type
Journal Article
Abstract
Seismic horizon tracking is fundamental to subsurface interpretation, yet it remains challenging in geologically complex settings characterized by faults, local structural uplift, and weak or discontinuous reflectors. Conventional auto-tracking methods rely heavily on seismic amplitudes and are therefore prone to drift, instability, and poor global consistency when reflector continuity is disrupted. We propose a confidence-aware framework jointly constrained by relative geologic time (RGT) and geologically informed prompt points for 2D and 3D seismic horizon interpretation. RGT is predicted from seismic data using an adversarially regularized network designed to retain structurally meaningful local variations and provide consistent stratigraphic guidance. Horizon extraction is then formulated as a prompt-driven local spreading process in which user-defined points supply geological control and overlapping RGT predictions guide stable propagation. Multiple prompts, including well ties and interpreter-defined control points, provide additional anchors for correcting long-range drift and improving within-domain tracking consistency. We further introduce a conditional tracking-confidence assessment by perturbing local prompt positions and measuring the dispersion among the resulting RGT-guided tracking solutions. Low dispersion indicates stable propagation, whereas high dispersion identifies regions that are sensitive to prompt initialization and may require additional geological control. Experiments using synthetic data and field data from the F3 and Opunake surveys demonstrate coherent tracking across structurally deformed, weak-reflection, and fault-dominated regions. The proposed framework integrates automated propagation, sparse expert constraints, and interpretable stability diagnostics into a unified semi-automatic workflow that identifies stable horizon segments and prioritizes prompt-sensitive regions for refinement.
Date Issued
2026-09-01
Date Acceptance
2026-08-20
Citation
Applied Computing and Geosciences, 2026, 32
ISSN
2590-1974
Publisher
Elsevier
Journal / Book Title
Applied Computing and Geosciences
Volume
32
Copyright Statement
© 2026 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Identifier
10.1016/j.acags.2026.100400
Publication Status
Published
Article Number
100400
Date Publish Online
2026-09-09