Active Learning with Deep Autoencoders for Seismic Facies Interpretation

Ahmad Mustafa, Ghassan AlRegib

TL;DR

Researchers have developed a "first-of-a-kind" active learning framework that uses deep autoencoders to drastically reduce the amount of expert-labeled data needed for seismic facies interpretation. By identifying the most "informative" seismic slices for a geologist to label—specifically those the model struggles to reconstruct—the system achieves a 10% accuracy boost over standard methods using only a handful of annotated sections.

What Problem Does This Paper Address?

In the world of oil and gas exploration, "seismic facies interpretation" is the process of partitioning 3D seismic volumes into different rock types (facies). Traditionally, this is a manual, Herculean task performed by expert geologists. While Deep Learning (DL) can automate this, it suffers from a "labeling bottleneck": DL models require massive amounts of pixel-perfect, human-annotated data to be accurate.

In geophysics, expert time is expensive. Most existing "data-efficient" methods assume you already have a labeled dataset and just want to use it better. This paper tackles the problem from the other end: Active Learning (AL). Instead of labeling everything upfront, the model "asks" the expert to label only the most useful samples in real-time, maximizing performance while minimizing the human workload.

Methods

The authors propose a unique framework based on Manifold Learning using Deep Autoencoders.

  1. Joint Learning Architecture: They use an encoder-decoder network that performs two tasks simultaneously:
  2. Unsupervised: Reconstructing the input seismic image (learning the "manifold" or the underlying structure of the data).
  3. Supervised: Segmenting the image into facies.
  4. The Query Strategy: This is the "active" part. After a training cycle, the model scans the unlabeled seismic volume. It looks for slices with the highest reconstruction error. The logic is simple: if the autoencoder can’t accurately "redraw" a slice, it means that slice contains features the model hasn't learned yet.
  5. Human-in-the-Loop: The system presents these high-error slices to the expert (the "Oracle") for labeling. These new labels are added to the training set, and the cycle repeats.
  6. Warm-Starting: Rather than retraining from scratch each time new labels are added, the authors use "curriculum learning" (warm-starting from the previous cycle's weights), which they found significantly improves convergence and final accuracy.

Datasets

The study utilizes the Netherlands offshore F3 block, a standard benchmark in the geophysics community.

  • Scope: A 3D volume consisting of 901 crosslines and 601 inlines.
  • Labels: Six major lithostratigraphic units (rock facies) interpreted from borehole data and seismic features.
  • Setup: The authors used a 3-fold cross-validation scheme, splitting the volume into training/validation and hold-out test sets to ensure the results weren't just a fluke of a specific data split.

Key Insights and Ideas

  • Reconstruction as a Proxy for Uncertainty: Most active learning uses "entropy" (how confused the model is about a label). This paper shows that reconstruction error is a superior metric for seismic data. If the model doesn't understand the "texture" of the rocks (the manifold), it won't be able to label them correctly.
  • The Power of Joint Tasks: By forcing the network to learn how to reconstruct the data and classify it at the same time, the internal features become much more robust. The unsupervised task "anchors" the supervised task.
  • Efficiency is King: The model reaches high performance levels with less than 1% of the total available data labeled. This suggests that much of a 3D seismic volume is redundant; you only need to label the "difficult" parts.

Key Results

  • Accuracy Boost: The proposed method achieved nearly a 10 percentage point difference in mean Intersection over Union (mIOU) compared to a baseline of random sampling.
  • Minimal Labeling: With only three fully annotated seismic sections, the active learning approach significantly outperformed standard training methods.
  • Qualitative Superiority: Visual inspections of the model's predictions (Figures 5 and 6 in the paper) show that the active learning model captures thin, under-represented facies layers that random sampling completely misses.

Limitations and Future Work

While the results are promising, a few areas remain for exploration:

  • Computational Overhead: Scanning an entire 3D volume for reconstruction errors after every training cycle can be computationally expensive as volumes grow to terabyte scales.
  • Expert Availability: The "human-in-the-loop" assumes an expert is standing by to label slices instantly. In a real-world workflow, the latency of human intervention might be a bottleneck.
  • Generalization: The authors suggest this framework could be applied to other tasks like salt body delineation or fault detection, but this study focused specifically on facies. Future work could test if the "reconstruction error" strategy holds up across different types of geophysical anomalies.

Summary updated

Abstract

Machine learning-assisted seismic interpretation tasks require large quantities of labeled data annotated by expert interpreters, which is a costly and time-consuming process. Where existing works to minimize dependence on labeled data assume the data annotation process to already be completed, active learning---a field of machine learning---works by selecting the most important training samples for the interpreter to annotate in real time simultaneously with the training of the interpretation model itself, resulting in high levels of performance with fewer labeled data samples than otherwise possible. Where active learning has been significantly performed for classification tasks with respect to natural images, there exist very little to no works for dense prediction tasks in geophysics like interpretation. We develop a unique and first-of-a-kind active learning framework for seismic facies interpretation using the manifold learning properties of deep autoencoders. By jointly learning representations for supervised and unsupervised tasks and then ranking unlabeled samples by their nearness to the data manifold, we are able to identify the most relevant training samples to be labeled by the interpreter in each training. On the popular F3 dataset, we obtain close to 10 percentage point difference in terms of interpretation accuracy between the proposed method and the baseline with only three fully annotated seismic sections.
BibTeX
@inproceedings{mustafa2023active,
  author = {Ahmad Mustafa and Ghassan AlRegib},
  title = {{Active Learning with Deep Autoencoders for Seismic Facies Interpretation}},
  booktitle = {Geophysics},
  year = {2023},
  doi = {10.1190/geo2022-0353.1},
}
Figure