Skip to main contentSkip to main content
Stage 3

DTW + GCN Horizon & Fault Tracking

Replace weeks of manual horizon picking with automated, uncertainty-quantified tracking across full 3D seismic volumes — with physical stratigraphic ordering enforced as a hard constraint.

The Problem

Manual horizon picking on a full 3D seismic cube can take months of interpreter time. Conventional seed-and-grow methods fail across faults, noise bursts, and amplitude dimming zones, requiring constant manual intervention and producing inconsistent results between interpreters.

The Solution

Dynamic Time Warping with a Sakoe-Chiba band constraint aligns trace envelopes globally, bridging faults and noise gaps that defeat local correlation. A Graph Convolutional Network propagates fault seeds through a 3D adjacency graph to complete fault planes. Per-voxel MC Dropout uncertainty maps flag low-confidence regions for targeted human review.

Technical Specifications

Implementation details

ComponentMethodNotes
Horizon TrackingDynamic Time WarpingSakoe & Chiba (1978). Sakoe-Chiba band constraint limits warping path width, preventing degenerate alignments on dipping reflectors. DOI
Fault PropagationGraph Convolutional NetworkKipf & Welling (2017). Semi-supervised GCN propagates fault seed points across adjacency graph of seismic voxels. DOI
Multi-class OutputSoft-argmax + temperature scalingNibali et al. (2018) soft-argmax for differentiable peak localisation; Guo et al. (2017) temperature scaling for calibrated confidence. DOI
Post-processingStratigraphic ordering DPHalpert (2018). Dynamic programming enforces physical stratigraphy — no horizon crossing — as a hard constraint.
Test-time AugmentationTTA 8-flip ensembleWang et al. (2019). Prediction averaged over all axis-aligned 3D flips reduces variance and improves boundary sharpness.
Uncertainty QuantificationMC DropoutGal & Ghahramani (2016). T=30 stochastic forward passes with Dropout3d active; per-voxel variance maps exported alongside predictions. DOI

Quality Gates

Automated acceptance criteria

Tracked horizons and faults are evaluated against held-out test labels before advancing. Low-confidence predictions are flagged automatically for human-in-the-loop review.

IoU> 0.6

Intersection-over-Union on held-out test horizons. Wu (2019) tolerant IoU variant accommodates sub-sample localisation error.

Boundary F1> 0.5

Saito & Rehmsmeier (2015). F1 score computed on dilated boundary masks; penalises horizon smearing more than interior mis-classification.

Tolerant IoUInformational

Wu et al. (2019). Fault-bench tolerant IoU inflates predicted and ground-truth masks before computing overlap, appropriate for thin fault planes.

Uncertainty-Aware Interpretation

Per-voxel confidence maps

MC Dropout (Gal & Ghahramani, 2016) runs T=30 stochastic forward passes with Dropout3d active. The variance across passes forms a per-voxel epistemic uncertainty map. Regions above the uncertainty threshold are automatically forwarded to the HITL Review stage — so geoscientists focus their attention where the model is least certain.

BALD (Houlsby et al., 2011) and entropy-based acquisition functions are computed alongside variance, enabling active learning workflows that iteratively improve the model on the hardest examples from your specific survey.

Stochastic passesT = 30
Uncertainty metricEpistemic variance
Acquisition fnBALD + Entropy
UQ output formatPer-voxel float32

References

Peer-reviewed foundations

  1. 1Sakoe, H. & Chiba, S. (1978). Dynamic Programming Algorithm Optimization for Spoken Word Recognition. IEEE TASSP 26(1). DOI
  2. 2Kipf, T. N. & Welling, M. (2017). Semi-Supervised Classification with Graph Convolutional Networks. ICLR. DOI
  3. 3Nibali, A. et al. (2018). Numerical Coordinate Regression with Convolutional Neural Networks. arXiv:1801.07372. DOI
  4. 4Guo, C. et al. (2017). On Calibration of Modern Neural Networks. ICML. DOI
  5. 5Gal, Y. & Ghahramani, Z. (2016). Dropout as a Bayesian Approximation. ICML. DOI
  6. 6Saito, T. & Rehmsmeier, M. (2015). The Precision-Recall Plot is More Informative than the ROC Plot. PLOS ONE. DOI
  7. 7Wu, X. et al. (2019). FaultSeg3D: Using Synthetic Data Sets to Train an End-to-End Convolutional Neural Network for 3D Seismic Fault Segmentation. Geophysics 84(3). DOI
  8. 8Halpert, A. (2018). Stratigraphically Ordered Horizon Tracking. SEG Technical Program Expanded Abstracts.
Stage 3 — Autotracking

Autotrack your survey

From a single well-tie seed point to a fully tracked 3D horizon grid — in hours, not months.