A dataset and benchmark of carbonate thin-section images for deep learning

Area of Science:

• Geoscience
• Computer Science
• Petroleum Geology

Background:

• Deep learning models are increasingly used for analyzing carbonate thin-section images.
• A significant limitation in this field is the absence of large, publicly available datasets, hindering reproducibility and model comparison.

Purpose of the Study:

• To introduce DeepCarbonate, a cleaned and standardized benchmark dataset for carbonate thin-section image analysis.
• To facilitate reproducible research and fair evaluation of deep learning models in carbonate petrography.

Main Methods:

• Collected carbonate samples from diverse geological formations (Ediacaran, Cambrian, Triassic, Cretaceous) in China and UAE.
• Curated the dataset with petroleum geology experts, removing invalid images via expert voting and 2σ filtering.
• Organized images in ImageNet format, creating 22 lithological categories hierarchically structured by optical mode (PPL, XPL, R) into train, validation, and test subsets.

Main Results:

• Evaluated multiple deep learning models (ResNet, VGG, DenseNet, MobileNet, EfficientNet) using PyTorch with CUDA acceleration.
• Conducted experiments under various conditions, including baseline, ablation, long-tailed distribution, and balanced Top 9 subsets.
• Demonstrated the dataset's effectiveness as a robust benchmark for carbonate petrography.

Conclusions:

• DeepCarbonate addresses the critical need for standardized datasets in deep learning for carbonate image analysis.
• The dataset supports reproducible research and facilitates the development and comparison of advanced machine learning models in the field.
• This work provides a valuable resource for advancing carbonate petrography research and its practical applications.