Seismic Facies Analysis: A Deep Domain Adaptation Approach

 

Abstract: This article presents a study on the application of deep neural networks (DNNs) for the accurate interpretation and classification of seismic facies in the field of geoscience. The study addresses the challenge of limited labeled data and distribution shifts between source and target domains. To overcome these challenges, the authors propose a deep neural network architecture called EarthAdaptNet (EAN) and introduce the use of unsupervised Deep Domain Adaptation (DDA) techniques.

The experiments conducted in this study utilize seismic images from the F3 block 3D dataset (source domain) located offshore Netherlands and the Penobscot 3D survey data (target domain) from Canada. The seismic images contain three geological classes with similar reflection patterns. The EAN architecture incorporates a transposed residual unit in the decoder block to enhance semantic segmentation of seismic images, particularly for classes with limited labeled data.

The proposed EAN architecture achieves a pixel-level accuracy of over 84% and an accuracy of approximately 70% for minority classes, demonstrating improved performance compared to existing architectures. Furthermore, the authors introduce the CORAL (Correlation Alignment) method to the EAN architecture, creating the EAN-DDA network. This network facilitates the classification of seismic reflections from the F3 and Penobscot datasets, showcasing a potential approach in scenarios where labeled data are unavailable. The EAN-DDA achieves a maximum class accuracy of approximately 99% for class 2 of the Penobscot dataset, with an overall accuracy exceeding 50%.

Index Terms: CORAL, Deep Learning, Domain Adaptation, EarthAdaptNet, Seismic Facies, Semantic Segmentation.

 

I. INTRODUCTION: Accurate interpretation of geologic features and reservoir properties is crucial for successful hydrocarbon exploration and production. The automation of seismic interpretation using deep neural networks (DNNs) has gained significant interest in recent years. However, the scarcity of large annotated datasets for seismic interpretation poses a challenge. To address this challenge, various techniques such as weakly-supervised learning, similarity-based data retrieval, and weakly-supervised label mapping algorithms have been proposed. Unsupervised machine learning techniques and alternative network architectures have also been explored.

Transfer learning has shown promise in reducing the costs associated with training DNNs from scratch. However, its application in domains such as medical imaging and earth science is challenging due to the requirement of large annotated datasets. Moreover, in the field of earth science, domain shifts are common, making the generalization of trained models to new datasets a complex task. Unsupervised Deep Domain Adaptation (DDA) methods have been proposed to address these challenges by allowing knowledge transfer across domains without relying on target domain labels.

This study focuses on the transfer of domain knowledge between different stratigraphic locations using reflection pattern similarities observed in seismic images. The authors propose the EarthAdaptNet (EAN) architecture, which leverages Residual Blocks (RBs) and Transposed Residual Blocks (TRBs) to address the issue of vanishing gradients. Additionally, they introduce the concept of DDA by incorporating the CORAL method, which minimizes the difference between source and target correlations.

The contributions of this article include the proposal of the EAN architecture for accurate seismic facies classification, especially for classes with limited labeled data. The integration of the CORAL method into EAN enables unsupervised deep domain adaptation. The performance of the proposed EAN-DDA network is evaluated in a multiclass classification problem for seismic facies analysis.

The remainder of the article is organized as follows: Section 2 provides an overview of the network architectures and the proposed approach. Subsequent sections present the experimental setup, results, and discussions. Finally, conclusions and future research directions are presented.

 

II. NETWORK ARCHITECTURES AND PROPOSED APPROACH: This section provides a summary of the network architectures used in the study and introduces the proposed approach for seismic facies analysis.

The EarthAdaptNet (EAN) architecture is designed to achieve accurate semantic segmentation of seismic images, particularly for classes with limited labeled data. The architecture incorporates Residual Blocks (RBs) and Transposed Residual Blocks (TRBs) with skip connections to address the issue of vanishing gradients. Additionally, the CORAL (Correlation Alignment for Domain Adaptation) method is introduced to facilitate unsupervised deep domain adaptation.

 

III. EXPERIMENTAL SETUP: In this section, the seismic image datasets from the F3 block (source domain) and the Penobscot survey (target domain) are described. The datasets contain three representative seismic facies classes with similar depositional and compositional environments. The authors follow the standard protocol of domain adaptation and utilize all labeled source domain data and unlabeled target domain data. Patch images are generated for the domain adaptation facies classification problem.

 

IV. RESULTS AND DISCUSSIONS: The performance of the proposed EarthAdaptNet (EAN) and EarthAdaptNet with Deep Domain Adaptation (EAN-DDA) architectures is evaluated in terms of pixel-level accuracy and class accuracy for seismic facies classification. The results demonstrate that the EAN architecture achieves higher performance compared to baseline architectures, particularly for minority classes with limited labeled data. The integration of the CORAL method in the EAN architecture further improves the classification accuracy for the target domain seismic facies.

 

V. CONCLUSIONS: In this section, the main contributions of the study are summarized. The proposed EarthAdaptNet (EAN) architecture demonstrates accurate semantic segmentation of seismic facies, outperforming existing architectures, especially for classes with limited labeled data. The integration of the CORAL method enables unsupervised deep domain adaptation, showcasing the potential for accurate classification of seismic facies in target domains. Future research directions and potential applications of the proposed approach are also discussed.

Overall, this article presents a comprehensive study on the application of deep neural networks and domain adaptation techniques for seismic facies analysis. The proposed EarthAdaptNet architecture and the incorporation of the CORAL method demonstrate promising results in accurately classifying seismic facies, even in scenarios with limited labeled data and distribution shifts between domains.

 

Full Research Paper Link: The full research paper can be accessed at the following link: Seismic Facies Analysis: A Deep Domain Adaptation Approach