Computer vision for carbonate core classification
File(s)
Author(s)
Dawson, Harriet
Type
Thesis
Abstract
Core data is a fundamentally important resource in characterising the geological subsurface. Globally, cores provide a record of diverse past geologic, climatic, and biological processes that have operated over multiple timescales. Leveraging insights from core images presents a pivotal challenge due to the largely subjective nature of traditional core description methods, inherent complexities and variability in geological data, and the paucity of significantly large, labelled datasets. Deep learning presents a data-driven approach to deriving predictive models from observational data.
With the evolution of machine learning, and growing awareness of the importance of big data, new automated workflows using deep learning algorithms present alternative approaches to the extensive time and labour requirements of traditional, manual core descriptions. Focusing solely on carbonate rocks, this thesis aims to establish foundations for applying deep learning to core interpretation. First, convolutional neural networks (CNNs) are used to classify carbonate textures from core images, optimising a workflow for lithological prediction across datasets of different magnitudes. To address the limiting factor of insufficient data, Generative Adversarial Networks (GANs) are applied to generate synthetic core images, thus filling sampling gaps and reducing overfitting. The performance of synthetic data is then compared with traditional data augmentation methods. Lastly, CNN-based object detection algorithms are applied to identify and quantify multiple types of carbonate grain, with results benchmarked against human interpretations to evaluate the efficacy of the method.
This approach uses computer vision models, such as CNNs and GANs, to enable streamlined geological interpretation and data-driven insights. As geoscience data availability increases through new datasets and the digitalisation of historical records, these methods can be further developed. The frameworks proposed in this thesis offer a pathway for advancing deep learning-based carbonate classification and have the potential for broader application to cores from diverse formations and lithologies.
With the evolution of machine learning, and growing awareness of the importance of big data, new automated workflows using deep learning algorithms present alternative approaches to the extensive time and labour requirements of traditional, manual core descriptions. Focusing solely on carbonate rocks, this thesis aims to establish foundations for applying deep learning to core interpretation. First, convolutional neural networks (CNNs) are used to classify carbonate textures from core images, optimising a workflow for lithological prediction across datasets of different magnitudes. To address the limiting factor of insufficient data, Generative Adversarial Networks (GANs) are applied to generate synthetic core images, thus filling sampling gaps and reducing overfitting. The performance of synthetic data is then compared with traditional data augmentation methods. Lastly, CNN-based object detection algorithms are applied to identify and quantify multiple types of carbonate grain, with results benchmarked against human interpretations to evaluate the efficacy of the method.
This approach uses computer vision models, such as CNNs and GANs, to enable streamlined geological interpretation and data-driven insights. As geoscience data availability increases through new datasets and the digitalisation of historical records, these methods can be further developed. The frameworks proposed in this thesis offer a pathway for advancing deep learning-based carbonate classification and have the potential for broader application to cores from diverse formations and lithologies.
Version
Open Access
Date Issued
2024-03-01
Date Awarded
01/02/2025
License URL
Advisor
John, Cédric M.
Dubrule, Olivier
Publisher Department
Department of Earth Science & Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)