Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

High-Resolution Detection of Earth Structural Heterogeneities from Seismic Amplitudes using Convolutional Neural Networks with Attention layers

Luiz Schirmer, Guilherme Schardong, Vinícius da Silva, Rogério Santos, Hélio Lopes

TL;DR

The paper tackles automated detection of seismic structural heterogeneities (faults/fractures) under limited labeled data. It introduces a lightweight, fully convolutional CNN augmented with attention blocks (SE-Net or self-attention) and uses transfer learning from synthetic data to real seismic patches, achieving superior IoU and precision with around half the parameters of competitors. Self-attention yields the best performance on the F3 North Sea dataset (IoU 0.912, precision 0.957) and generalizes to the Great South Basin (IoU 0.86, F1 0.88), demonstrating data-efficient, scalable detection. The approach offers a practical, fast, and low-cost tool for automated interpretation in hydrocarbon exploration and production decisions.

Abstract

Earth structural heterogeneities have a remarkable role in the petroleum economy for both exploration and production projects. Automatic detection of detailed structural heterogeneities is challenging when considering modern machine learning techniques like deep neural networks. Typically, these techniques can be an excellent tool for assisted interpretation of such heterogeneities, but it heavily depends on the amount of data to be trained. We propose an efficient and cost-effective architecture for detecting seismic structural heterogeneities using Convolutional Neural Networks (CNNs) combined with Attention layers. The attention mechanism reduces costs and enhances accuracy, even in cases with relatively noisy data. Our model has half the parameters compared to the state-of-the-art, and it outperforms previous methods in terms of Intersection over Union (IoU) by 0.6% and precision by 0.4%. By leveraging synthetic data, we apply transfer learning to train and fine-tune the model, addressing the challenge of limited annotated data availability.

High-Resolution Detection of Earth Structural Heterogeneities from Seismic Amplitudes using Convolutional Neural Networks with Attention layers

TL;DR

The paper tackles automated detection of seismic structural heterogeneities (faults/fractures) under limited labeled data. It introduces a lightweight, fully convolutional CNN augmented with attention blocks (SE-Net or self-attention) and uses transfer learning from synthetic data to real seismic patches, achieving superior IoU and precision with around half the parameters of competitors. Self-attention yields the best performance on the F3 North Sea dataset (IoU 0.912, precision 0.957) and generalizes to the Great South Basin (IoU 0.86, F1 0.88), demonstrating data-efficient, scalable detection. The approach offers a practical, fast, and low-cost tool for automated interpretation in hydrocarbon exploration and production decisions.

Abstract

Earth structural heterogeneities have a remarkable role in the petroleum economy for both exploration and production projects. Automatic detection of detailed structural heterogeneities is challenging when considering modern machine learning techniques like deep neural networks. Typically, these techniques can be an excellent tool for assisted interpretation of such heterogeneities, but it heavily depends on the amount of data to be trained. We propose an efficient and cost-effective architecture for detecting seismic structural heterogeneities using Convolutional Neural Networks (CNNs) combined with Attention layers. The attention mechanism reduces costs and enhances accuracy, even in cases with relatively noisy data. Our model has half the parameters compared to the state-of-the-art, and it outperforms previous methods in terms of Intersection over Union (IoU) by 0.6% and precision by 0.4%. By leveraging synthetic data, we apply transfer learning to train and fine-tune the model, addressing the challenge of limited annotated data availability.
Paper Structure (11 sections, 6 equations, 9 figures, 1 table)

This paper contains 11 sections, 6 equations, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Geoscientific Background
  4. Methods and Network Architecture
  5. Convolutional model and synthetic dataset
  6. Attention block
  7. Squeeze-and-excitation networks
  8. Self-attention
  9. Transfer Learning Approach
  10. Experiments
  11. Conclusion

Figures (9)

  • Figure 1: Seismic Inline 120 within TWT (ms) showing the base of Cenozoic reflection (yellow line) in F3 Block. A gas chimney adjacent to a fault, and immediately overlying an apparently leaking bright spot at reservoir levels. A seismic pull-down effect can be seen underneath the chimney. Green segments can be modeled as faults
  • Figure 2: Seismic Inline 220 within TWT (ms) showing the base of Cenozoic reflection (yellow line). Leakage suggestions are indicated as fault systems in the northern part of F3 block. Faults provide gas migration paths and control bright spot indicators. All green segments can be modeled as faults.
  • Figure 3: Seismic Inline 690 within TWT (ms) showing the base of Cenozoic reflection (yellow line) in F3 Block. A shallow bright spot over a flat spot corresponds to the gas-water contact in the Upper Pliocene sediments. This seismic expression is indicative of effective structural gas trapping. Green segments could be modeled as faults
  • Figure 4: The image at the top is a synthetic image generated using the IPF code developed by HaleHale2014. Below it, we have an image with a dashed line representing fault data annotations. During training, these annotations are converted into a binary mask.
  • Figure 5: Architecture of our neural network. Our model is based on a fully convolutional architecture where similarly to DenseNets we have a residual operation before the attention block. We generate two versions of this architecture, one using SE-Nets and the other using a Self-attention block.
  • ...and 4 more figures