Related Experiment Video
Updated: Aug 22, 2025
12:18
Pore-scale Imaging and Characterization of Hydrocarbon Reservoir Rock Wettability at Subsurface Conditions Using X-ray Microtomography
Published on: October 21, 2018
14.1K
Segmentation for Multi-Rock Types on Digital Outcrop Photographs Using Deep Learning Techniques
Owais A Malik1,2, Idrus Puasa3, Daphne Teck Ching Lai1,2
1School of Digital Science, Universiti Brunei Darussalam, Brunei Darussalam, Gadong BE1410, Brunei.
Sensors (Basel, Switzerland)
|November 11, 2022
Summary
This study introduces machine learning for identifying sandstone and mudstone in rock images, automating sedimentology. Semantic segmentation models like U-Net and LinkNet achieved high accuracy, showing potential for reservoir characterization.
Area of Science:
- Geosciences and Artificial Intelligence
- Application of machine learning in sedimentology and reservoir characterization.
Background:
- Manual identification and classification of sedimentary rocks (sandstone, mudstone) are time-consuming and challenging in field studies.
- Existing machine learning approaches often fail to classify multiple rock types within a single image or use artificial datasets.
Purpose of the Study:
- To test the efficacy of machine learning, specifically semantic segmentation, for multi-rock identification using high-resolution photographs.
- To evaluate the performance of U-Net and LinkNet models with various backbones and image processing techniques for rock classification.
Main Methods:
- Applied U-Net and LinkNet semantic segmentation models to a dataset of 102 field images of sandstone and mudstone.
- Utilized four pre-trained convolutional neural networks (Resnet34, Inceptionv3, VGG16, Efficientnetb7) as backbones for the segmentation models.
- Investigated the impact of image enhancement and different color representations (e.g., L*a*b*) on model performance.
Main Results:
- The ensemble of U-Net models achieved the highest performance with a mean over intersection (MIoU) of 0.8201.
- LinkNet with Efficientnetb7 backbone showed strong individual performance (MIoU = 0.8135).
- U-Net with Efficientnetb7 using L*a*b* color representation yielded the best results for segmenting individual rock types (MIoU = 0.8178).
Conclusions:
- Semantic segmentation models, particularly U-Net, demonstrate significant potential for automating the multi-rock classification process in sedimentology.
- This automated approach can streamline reservoir characterization by enabling efficient extraction of rock data for further analysis and modeling.
