Related Experiment Video
Updated: Sep 17, 2025
10:25
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
9.4K
Comparative analysis of sandstone microtomographic image segmentation using advanced convolutional neural networks
Mazaher Hayatdavoudi1, Mohammad Emami Niri2, Ahmad Kalhor3
1Institute of Petroleum Engineering, School of Chemical Engineering, College of Engineering, University of Tehran, Tehran, Iran.
Scientific Reports
|July 2, 2025
Summary
Deep learning models, specifically Convolutional Neural Networks (CNNs), significantly improve digital rock image segmentation for enhanced reservoir characterization. These advanced CNNs outperform traditional methods in accuracy and predicting fluid flow properties.
Area of Science:
- Geosciences and Petroleum Engineering
- Artificial Intelligence and Machine Learning
Background:
- Automated segmentation of digital rock images is crucial for reservoir characterization, impacting porosity and fluid flow evaluations.
- Traditional methods like Otsu thresholding have limitations in accurately segmenting complex pore structures.
Purpose of the Study:
- To explore and benchmark state-of-the-art Convolutional Neural Network (CNN) architectures for segmenting rock micro-CT images.
- To evaluate the performance of various CNN models against traditional methods for enhancing reservoir characterization efficiency.
Main Methods:
- Implementation and comparison of diverse CNN architectures (e.g., Fully Convolutional Networks, Encoder-Decoder, Attention-Based Models).
- Benchmarking CNN segmentation accuracy against the Otsu thresholding method using a dataset of 5,000 2D sandstone slices.
- Utilizing pixel-wise accuracy metrics (F1-score, binary-IOU, Recall, Precision) for evaluation.
- Employing numerical simulation methods (LBM, PNM, CFD) to predict permeability and formation factor using CNN-segmented images.
Main Results:
- Advanced CNNs demonstrate superior pixel-wise segmentation accuracy compared to the Otsu method.
- CNNs achieve significantly better performance in predicting fluid flow characteristics like permeability and formation factor.
- EfficientNetB0-Unet, VGG16-Unet, and Enet showed exceptional results in segmenting complex pore structures.
Conclusions:
- Deep learning, particularly CNNs, offers a substantial advancement over traditional methods for digital rock image analysis.
- The study validates the effectiveness of specific CNN architectures for accurate pore structure segmentation and reliable prediction of petrophysical properties.
- Optimized CNN models enhance the efficiency and precision of reservoir characterization through improved image analysis and simulation inputs.
