zhangjw0110/drill-core-image-dataset

Dataset Card for Drill Core Image Dataset (DCID)

Dataset Details

Dataset Description

The Drill Core Image Dataset (DCID) is a large-scale benchmark designed for lithology classification based on RGB core images. It provides two primary versions:

• —DCID-7: 7 lithology categories with 5,000 images per class.
• —DCID-35: 35 lithology categories with 1,000 images per class.

All original images are 512×512 pixels in resolution. Each category is split into training and testing subsets in an 8:2 ratio. Additional variants are generated by resizing to smaller resolutions (32, 64, 128, 256) and applying real-world data augmentation (RWDA) to simulate image imperfections.

• —Curated by: Jia-Yu Li, Ji-Zhou Tang, et al.
• —Shared by: Jia-Yu Li (lijiayu1120@tongji.edu.cn), Ji-Zhou Tang (jeremytang@tongji.edu.cn)
• —License: CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0)

Visual Overview

DCID Naming Convention

The dataset naming follows the DCID-R-C-L-I format:

• —R: resolution (32, 64, 128, 256, 512)
• —C: number of categories (7 or 35)
• —L: RWDA level (0.0 – 0.4)
• —I: injection scope (N, T, E, A)

[image]

DCID-7 Dataset

The DCID-7 dataset contains 35,000 images (5,000 per category). Each class has 4,000 training and 1,000 testing images (8:2 ratio). This version is suitable for evaluating model upper-bound performance.

[image]

DCID-35 Dataset

The DCID-35 dataset contains 35,000 images (1,000 per category). Each class has 800 training and 200 testing images. This fine-grained version is designed to assess model generalization under complex conditions.

[image]

Dataset Sources

• —GitHub Repository: https://github.com/JiayuLi1120/drill-core-image-dataset
• —Hugging Face Dataset: https://huggingface.co/datasets/168sir/drill-core-image-dataset
• —Paper: https://doi.org/10.1016/j.petsci.2025.04.013

Usage

Step 1: Download and extract

Download the DCID.zip archive from Hugging Face and extract it:

unzip DCID.zip -d ./DCID

This will give you the following folders:

• —DCID-512-7/ and noise-512-7/
• —DCID-512-35/ and noise-512-35/

Step 2: Build custom dataset versions

We provide a script `build_dcid_dataset.py` to generate different dataset variants.

Example: Create a 32×32 resolution, 7 classes, 40% RWDA (train set only) dataset:

python build_dcid_dataset.py \
    --root ./DCID \
    --R 32 \
    --C 7 \
    --L 0.4 \
    --I T \
    --out_dir ./output

This generates a new dataset at:

./output/DCID-32-7-0.4-T/

Script Parameters

• —`R`: target resolution (32, 64, 128, 256)
• —`C`: number of categories (7 or 35)
• —`L`: RWDA level (0.0–0.4)
• —`I`: injection scope:

• —N: none
• —T: train set only
• —E: test set only
• —A: all (train + test)

Citation

If you use this dataset in your work, please cite:

@article{Li2025DCID,
  title = {A large-scale, high-quality dataset for lithology identification: Construction and applications},
  author = {Jia-Yu Li and Ji-Zhou Tang and Xian-Zheng Zhao and Bo Fan and Wen-Ya Jiang and Shun-Yao Song and Jian-Bing Li and Kai-Da Chen and Zheng-Guang Zhao},
  journal = {Petroleum Science},
  year = {2025},
  issn = {1995-8226},
  doi = {10.1016/j.petsci.2025.04.013}
}