DeepCarbonate: A Dataset and Benchmark of Carbonate Thin-Section Images for Deep Learning

Keran Li^a,1,2, Jinmin Song^a,1,*, Zhaoyi Zhang^a,1, Shan Ren^a,1, Yang Lan^b,1, Di Yang^a, Zhiwu Li^a, Shugen Liu^a, Chunqiao Yan^a, Xin Jin^a, Shaohai Yang^a, Jiaxin Guo^a

^aState Key Laboratory of Oil and Gas Reservoir Geology and Exploitation, Chengdu University of Technology, Chengdu 610059, China

^bSchool of Economics and Management, Beihang University, Beijing, 100191, China

¹Equal contribution

²Present address: State Key Laboratory of Critical Earth Material Cycling and Mineral Deposits, Frontiers Science Center for Critical Earth Material Cycling, School of Earth Sciences and Engineering, Nanjing University, Nanjing, 210023, China

^*Corresponding authors

---

33.2 GB · 55,786 Images · 22 Lithological Categories · 3 Optical Modes

📄 Paper • 💾 Dataset (Zenodo) • 🌐 Project Page • 📊 Benchmarks

---

📋 Table of Contents

• Overview
• Geological Background
• Dataset Statistics
• Installation
• Usage
• Benchmark Results
• Citation

---

🎯 Overview

DeepCarbonate is a rigorously curated dataset for automated carbonate petrography using deep learning. Unlike generic image datasets, it addresses the multimodal nature of geological thin-section analysis requiring specific optical conditions (plane-polarized, cross-polarized, and reflected light) to distinguish mineralogical features.

🔬 Key Features

• Expert-validated: 10 petroleum geologists (6 Ph.D., 4 M.Sc.) voted on lithological labels; images with >20% disagreement excluded
• Multimodal imaging: PPL, XPL, and Reflected light modes for comprehensive mineralogical characterization
• Standardized format: Organized in ImageNet-style hierarchy (train/val/test)
• Quality controlled: 345 invalid images removed via 2σ filtering (blur, brightness, artifacts) + expert inspection
• Geologically diverse: Covers microbialites, grainstones, evaporites, and reservoir features (pores, fractures)

---

🌍 Geological Background

Samples were collected from carbonate reservoirs spanning ~500 million years of geological history:

Basin	Formation	Age	Lithology Focus
Sichuan Basin, China	Dengying	Ediacaran	Microbial carbonates, dolostones
Sichuan Basin, China	Longwangmiao	Cambrian	Grainstones, oolites
Sichuan Basin, China	Leikoupo & Jialingjiang	Triassic	Evaporites, lagoonal deposits
UAE	Mishrif	Cretaceous	Reefal limestones, bioclastic carbonates
Others	Unkown	Unkown	Reefal limestones, bioclastic carbonates

🏔️ 22 Lithological Categories

Click to expand class definitions (Folk & Dunham classification)

Class ID	Category	Description	Typical Features
1	Microbial mudstone	Gelatinous EPS-rich matrix	Micrite, microbial textures
2	Bioclastic carbonates	Shell/coral fragments	Allochem-rich, variable sorting
3	Breccia carbonates	Angular >2mm fragments	Tectonic/collapse features
4	Coarse crystal dolostone	Recrystallized dolomite	Euhedral crystals >100μm
5	Coarse crystal limestone	Sparitic calcite	Diagenetic recrystallization
6	Evaporite	Halite/gypsum precipitation	Arid environment indicator
7	Finely crystal dolostone	Early diagenetic dolomite	<10μm crystals, anhydrite assoc.
8	Finely crystal limestone	Microcrystalline calcite	Hypersaline pore waters
9	Foam carbonates	3D interconnected pore network	Microbial, high porosity
10	Cemented fracture	Calcite/silica filled fractures	Reservoir connectivity features
11	Botryoidal dolostone	Grape-like textures	Microbial precipitates
12	Micritic dolostone	<4μm dolomite crystals	Restricted marine environments
13	Micritic limestone	<63μm calcite matrix	Low permeability seal rocks
14	Mid crystal dolostone	Intermediate crystallinity	10-100μm crystals
15	Mid crystal limestone	Diagenetic texture	Moderate porosity
16	Oncolite	Spherical concentric structures	Nucleus-coated grains
17	Oolite	<2mm concentric ooids	High-energy shallow marine
18	Siliceous carbonates	Chert/opal bearing	Silica diagenesis
19	Stromatolites	Laminated microbial mats	Layered cyanobacterial structures
20	Stylolite	Pressure solution seams	Irregular interlocking columns
21	Thrombolite	Clotted microbial structure	Irregular mesoscopic fabric
22	Pore	Primary/cemented voids	Reservoir quality indicator

---

📊 Dataset Statistics

Overall Distribution

Subset	Images	Percentage	Size
Train	39,070	70.0%	~23.2 GB
Validation	11,157	20.0%	~6.6 GB
Test	5,559	10.0%	~3.4 GB
Total	55,786	100%	33.2 GB

Detailed Class Statistics

Rank	Class	Total Images	Train	Val	Test	Optical Modes
1	Pore	6,432	4,502	1,286	644	PPL, XPL
2	Thrombolite	5,328	3,730	1,066	532	PPL, XPL
3	Microbial mudstone	4,728	3,310	946	472	PPL, XPL
4	Stromatolites	3,569	2,498	714	357	PPL, XPL
5	Micritic limestone	3,231	2,262	646	323	PPL, XPL, ARS
6	Bioclastic carbonates	3,031	2,122	606	303	PPL, XPL
7	Micritic dolostone	2,974	2,082	595	297	PPL, XPL, ARS
8	Siliceous carbonates	2,938	2,057	588	293	PPL, XPL, R
9	Cemented fracture	2,788	1,952	558	278	PPL, XPL, R
10	Foam carbonates	1,218	853	244	121	PPL, XPL
11	Breccia carbonates	1,123	786	225	112	PPL, XPL
12	Stylolite	917	642	183	92	PPL, XPL
13	Evaporite	575	403	115	57	PPL, XPL
14	Mid crystal dolostone	424	297	85	42	PPL, XPL, ARS
15	Botryoidal dolostone	252	176	50	26	PPL, XPL
16	Finely crystal dolostone	234	164	47	23	PPL, XPL
17	Oolite	150	105	30	15	PPL, XPL
18	Oncolite	129	90	26	13	PPL, XPL
19	Coarse crystal dolostone	112	78	22	12	PPL, XPL
20	Coarse crystal limestone	108	76	22	10	PPL, XPL
21	Mid crystal limestone	14	10	3	1	PPL, XPL
22	Finely crystal limestone	14	10	3	1	PPL, XPL

Note: PPL = Plane-Polarized Light, XPL = Cross-Polarized Light, R = Reflected Light, ARS = Alizarin Red Stained

---

⚙️ Installation

Prerequisites

• Hardware: NVIDIA GPU with ≥80GB VRAM (tested on A800) for full training; 16GB+ for inference
• Software: Python 3.8+, CUDA 11.3+

Environment Setup

# Clone repository
git clone https://github.com/KeranLi/DeepCarbonate.git
cd DeepCarbonate

# Create conda environment
conda create -n deepcarbonate python=3.9
conda activate deepcarbonate

# Install core dependencies
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Key Dependencies:

torch >= 1.11.0
torchvision >= 0.12.0
numpy >= 1.20.0
pillow >= 8.0.0
matplotlib >= 3.3.0
scikit-learn >= 0.24.0
tqdm >= 4.60.0
pandas >= 1.3.0
seaborn >= 0.11.0

---

🚀 Usage

1. Data Preparation

Download the dataset from Zenodo and organize as follows:

DeepCarbonate/
├── PPL/                    # Plane-Polarized Light (~18,595 images)
│   ├── train/
│   │   ├── class1/         # Microbial mudstone
│   │   ├── class2/         # Bioclastic carbonates
│   │   └── ...             # class3-22
│   ├── val/
│   └── test/
├── XPL/                    # Cross-Polarized Light (~18,595 images)
│   ├── train/
│   ├── val/
│   └── test/
└── R/                      # Reflected Light (~4,000 images)
    ├── train/
    ├── val/
    └── test/

Naming Convention: LithologyName.Index.jpg (e.g., Bioclastic carbonates.2.jpg, Pore.22.jpg)

2. Training

Full 22-Class Training (Baseline)

python main.py \
    --mode train \
    --model resnet50 \
    --data_path ./PPL \
    --epochs 50 \
    --batch_size 512 \
    --lr 1e-5 \
    --num_classes 22 \
    --early_stopping

Top-9 Balanced Subset (Recommended for Stable Training)

python main.py \
    --mode train \
    --model resnet50 \
    --data_path ./PPL \
    --classes 1,2,3,4,5,6,7,8,9 \
    --epochs 50 \
    --batch_size 256

Ablation: Single Polarization Mode

# Train with PPL only (expect performance drop)
python main.py --mode train --model mobilenet --data_path ./PPL --ablation_mode single_polar

# Train without Alizarin Red Stained (ARS) images
python main.py --mode train --model resnet18 --exclude_ars

3. Evaluation

# Generate confusion matrix and full metrics
python eval.py \
    --checkpoint experiments/best_model.pth \
    --data_path ./PPL/test \
    --num_classes 22 \
    --save_results ./results/

4. Class Activation Mapping (CAM) Visualization

python visualize_cam.py \
    --image path/to/image.jpg \
    --model resnet50 \
    --checkpoint experiments/best_model.pth

---

🏆 Benchmark Results

Main Results: 22-Class Classification

Model	Accuracy	Precision	Recall	F1-Score	AUC	Top-1 Acc	Top-5 Acc
MobileNet	0.68	0.68	0.67	0.67	0.67	0.52	0.81
DenseNet	0.64	0.64	0.63	0.64	0.64	0.55	0.82
VGG16	0.62	0.62	0.61	0.62	0.63	0.57	0.84
EfficientNet	0.63	0.62	0.62	0.62	0.62	0.52	0.81
ResNet18	0.62	0.62	0.63	0.63	0.62	0.56	0.81
ResNet50	0.56	0.57	0.55	0.57	0.56	0.58	0.85

Key Findings:

• MobileNet achieves best overall metrics (Acc/F1: 0.68) despite being lightweight
• ResNet50 shows highest Top-5 accuracy (85%) but lowest Top-1 (56%), indicating hierarchical confusion
• Long-tail effect: Severe performance degradation on tail classes (14-112 samples) vs head classes (6,000+ samples)

Top-9 Balanced Subset Results

Model	Accuracy	Precision	Recall	F1-Score	AUC	Top-1 Acc	Top-9 Acc
ResNet50	0.85	0.85	0.85	0.86	0.85	0.72	0.90
ResNet18	0.83	0.82	0.83	0.83	0.83	0.78	0.81
EfficientNet	0.75	0.72	0.72	0.73	0.74	0.74	0.84
DenseNet	0.73	0.73	0.73	0.74	0.74	0.74	0.86
MobileNet	0.72	0.72	0.71	0.70	0.72	0.72	0.85
VGG16	0.72	0.82	0.72	0.71	0.72	0.67	0.88

Ablation Studies

Configuration	MobileNet Acc	ResNet50 Acc	Finding
All Modes	0.68	0.56	Baseline (multimodal)
w/o ARS	0.70	0.58	Alizarin Red staining removal has minimal impact
PPL Only	0.51	0.49	Single polarization causes severe feature loss
XPL Only	0.52	0.50	Cross-polarized alone insufficient
R Only	0.48	0.46	Reflected light limited for mineral ID

---

🖼️ Data Quality Control

Image Quality Metrics (2σ Filtering)

Following strict geological imaging standards:

Metric	Threshold	Rejected
Brightness	μ ± 2σ	152 images (low brightness)
Sharpness	Laplacian variance	174 images (blur)
Blur	Gaussian detection	Motion artifacts
Artifacts	Noise detection	19 images (damaged)

Final Clean Dataset: 55,441 valid images (99.4% retention rate)

Microscopy Parameters

• Equipment: Nikon Eclipse E600POL
• Illumination: Köhler method, 12V 100W halogen, stabilized at 9.8V
• CCD: 14-bit, 0.63× C-mount
• Magnification: 10× (primary), 5× (overview)
• Optics: Field/aperture diaphragms at 90%/70% NA (transmitted), 60% (reflected)

---

📚 Citation

If you use DeepCarbonate in your research, please cite:

@article{li2026deepcarbonate,
  title={A dataset and benchmark of carbonate thin-section images for deep learning},
  author={Li, Keran and Song, Jinmin and Zhang, Zhaoyi and Ren, Shan and Lan, Yang and Yang, Di and Li, Zhiwu and Liu, Shugen and Yan, Chunqiao and Jin, Xin and Yang, Shaohai and Guo, Jiaxin},
  journal={Scientific Data},
  volume={X},
  pages={XXX--XXX},
  year={2026},
  publisher={Nature Publishing Group},
  doi={10.1038/s41597-026-06633-5}
}

Dataset Citation:

@dataset{li2025deepcarbonate,
  author={Li, Keran and Song, Jinmin and Zhang, Zhaoyi and Ren, Shan and Lan, Yang and Yang, Di and Li, Zhiwu and Liu, Shugen and Yan, Chunqiao and Jin, Xin and Yang, Shaohai and Guo, Jiaxin},
  title={DeepCarbonate: A benchmark dataset for carbonate thin-section image analysis},
  year={2025},
  publisher={Zenodo},
  doi={10.5281/zenodo.18061204}
}

---

⚠️ Usage Notes

1. Long-tail Distribution: The dataset exhibits severe class imbalance (6432 vs 14 images). For production use, we strongly recommend:

   • Using the Top-9 balanced subset for initial model development
   • Applying class-weighted loss functions or oversampling for full 22-class training

2. Optimal Imaging: For inference on new samples, capture both PPL and XPL images simultaneously. Reflected light (R) is supplementary for organic matter/bitumen analysis.

3. Magnification: All images captured at 10× magnification under standardized Köhler illumination.

4. Staining: Alizarin Red S staining available for subset; use for distinguishing calcite vs dolomite.

---

📂 Project Structure

DeepCarbonate/
├── 📂 data/                       # Dataset directory (download separately)
│   ├── PPL/                       # Plane-polarized light images
│   ├── XPL/                       # Cross-polarized light images
│   └── R/                         # Reflected light images
├── 📂 experiments/                # Training outputs
│   ├── models/                    # Saved checkpoints (.pth)
│   ├── logs/                      # TensorBoard logs
│   └── results/                   # Evaluation metrics
├── 📂 src/                        # Source code
│   ├── dataset.py                 # Custom Dataset class
│   ├── model.py                   # ResNet, DenseNet, EfficientNet, MobileNet
│   ├── trainer.py                 # Training loop with early stopping
│   └── eval.py                    # Metrics calculation
├── main.py                        # Entry point
├── requirements.txt               # Dependencies
├── LICENSE                        # CC BY-NC-ND 4.0 for data, MIT for code
└── README.md                      # This file

---

🤝 Contributing

We welcome contributions to expand the dataset, especially:

• Additional samples for underrepresented classes (Classes 14-22)
• New optical modes (fluorescence, cathodoluminescence)
• Segmentation masks for mineral phase analysis

📝 License

• Dataset: CC BY-NC-ND 4.0 (Non-commercial use only)
• Code: MIT License