Feature extraction and classification of digital rock images via pre-trained convolutional neural network and unsupervised machine learning
Understanding the microstructure of porous media is crucial in various fields—particularly in petroleum engineering, hydrogeology, and materials science—because it directly influences the properties of porous materials and the behavior of fluids within their pores. Traditional characterization metho...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | Machine Learning: Science and Technology |
| Subjects: | |
| Online Access: | https://doi.org/10.1088/2632-2153/adcf71 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Understanding the microstructure of porous media is crucial in various fields—particularly in petroleum engineering, hydrogeology, and materials science—because it directly influences the properties of porous materials and the behavior of fluids within their pores. Traditional characterization methods often struggle to capture the complex, heterogeneous micro-scale features of rock structures. To address this challenge, this study presents a novel approach for the classification and visualization of rock microstructure from micro-computed tomography images, leveraging pre-trained convolutional neural network (CNN) models (AlexNet, GoogLeNet, Inception v3 Net, ResNet, and DenseNet) combined with unsupervised machine learning (USML) techniques principal component analysis, multidimensional scaling, isometric mapping, t-distributed stochastic neighbor embedding (t-SNE), and uniform manifold approximation projection (UMAP)). Using pre-trained CNNs allows us to extract rich feature representations without the need for large, specialized training datasets, effectively capturing intricate patterns in the microstructures. The application of USML methods enables us to reduce dimensionality and uncover latent structures in the data without supervision. We tested the effectiveness of our method through three distinct case studies that include a wide variety of porous structures and found high classification accuracy using DenseNet and t-SNE or UMAP. Our approach successfully distinguishes similar rock samples that have been difficult to classify using conventional features such as porosity, specific surface area, and Euler characteristics, as demonstrated by silhouette score, Davies–Bouldin Index, and Caliński–Harabasz Index. To enhance the interpretability of the machine learning approach, we proposed a patch-based analysis to identify local characteristic textural patterns that contribute significantly to the classification of individual rock samples. By visualizing the spatial distribution of these patterns and quantifying their characteristics, we gained insights into the microstructural differences between rock samples, providing an effective tool for interpreting the classification results and understanding the underlying factors that differentiate various rock types. |
|---|---|
| ISSN: | 2632-2153 |