Spatiotemporal Modeling of Seismic Images for Acoustic Impedance Estimation (2006.15472v1)

Abstract: Seismic inversion refers to the process of estimating reservoir rock properties from seismic reflection data. Conventional and machine learning-based inversion workflows usually work in a trace-by-trace fashion on seismic data, utilizing little to no information from the spatial structure of seismic images. We propose a deep learning-based seismic inversion workflow that models each seismic trace not only temporally but also spatially. This utilizes information-relatedness in seismic traces in depth and spatial directions to make efficient rock property estimations. We empirically compare our proposed workflow with some other sequence modeling-based neural networks that model seismic data only temporally. Our results on the SEAM dataset demonstrate that, compared to the other architectures used in the study, the proposed workflow is able to achieve the best performance, with an average $r^{2}$ coefficient of 79.77\%.

• The paper introduces a deep learning method that integrates spatial context with temporal modeling using 2-D TCNs for seismic inversion.
• It achieves notable performance with an average r² of 79.77% and a Pearson correlation of 0.9259, surpassing traditional 1-D models.
• The study demonstrates improved subsurface feature delineation, offering enhanced acoustic impedance estimation for effective reservoir exploration.

Spatiotemporal Modeling for Seismic Inversion: A Deep Learning Approach

The paper authored by A. Mustafa, M. Alfarraj, and G. AlRegib presents a novel method for seismic inversion by leveraging deep learning techniques to model seismic data not just temporally but also spatially. The specific focus is on estimating acoustic impedance (AI) from seismic reflection data, an essential process for determining reservoir rock properties crucial for oil and gas exploration.

Background and Motivation

Traditional seismic inversion techniques rely on trace-by-trace analysis, considering only temporal aspects of seismic data. This conventional approach inherently overlooks the spatial correlations present in seismic images, potentially leading to lateral discontinuities in property estimation. A notable challenge in the domain is utilizing the spatial structure of seismic data effectively, which is crucial given the highly correlated nature of neighboring seismic traces.

Methodology

The paper introduces a deep learning-based workflow employing a Temporal Convolutional Network (TCN) that extends conventional sequence modeling to incorporate spatial information in seismic images. The proposed network processes seismic data segments as rectangular patches, integrating spatial context with temporal modeling. This methodology contrasts with existing sequence modeling approaches that treat each seismic trace independently.

The architecture utilizes 2-D TCNs, drawing from sequence modeling works such as those by Bai et. al (2018). The feature extractor block includes multiple temporal blocks utilizing 2-D convolutions, allowing the network to capture temporal dependencies within traces while simultaneously leveraging spatial context. The network performance was quantitatively benchmarked using metrics such as the average $r^2$ coefficient and Pearson's Correlation Coefficient (PCC).

Results and Evaluation

The proposed spatiotemporal model was evaluated on the SEAM dataset, showing superior performance in estimating AI compared to conventional 1-D TCN and LSTM-based models. Notably, the 2-D TCN-based architecture achieved an average $r^2$ coefficient of 79.77% and PCC of 0.9259, outpacing the compared sequence models. The empirical results highlighted the enhanced delineation of subsurface features, including salt regions, demonstrating the network's ability to better capture the vertical and horizontal seismic attributes.

Discussion and Implications

The implications of this paper are twofold:

1. Practical Implications: The ability to integrate spatial and temporal information provides a more comprehensive tool for seismic inversion, resulting in more accurate subsurface models. This has significant implications for improving geophysical surveys in the energy sector, particularly for projects with sparse data.
2. Theoretical Implications: By successfully extending convolutional architectures to encompass spatial dimensions in image-like seismic data, the paper suggests potential for further exploration in applying such hybrid models to other domains involving spatiotemporal data.

Future Directions

Future research could investigate more complex 3D models and explore the integration of other machine learning architectures that focus on capturing deeper spatial relationships. Additionally, the framework might benefit from incorporating other geological constraints and priors directly into the model design, potentially enhancing its robustness to noisy data. As deep learning continues to advance, hybrid approaches that combine the strengths of various neural architectures could see increasing exploration within seismic data processing paradigms.

In conclusion, this paper presents a substantial advancement in leveraging AI for seismic inversion by treating the data as spatiotemporal entities, thus opening pathways for enhanced modeling in geophysical exploration.