BiliSakura/JL1-CUP-2024-Second-Format

JL1 CUP 2024 — Second-track format for semantic change detection

Bi-temporal 256×256 RGB patches with per-pixel semantic maps at times T1/T2 and a binary change map, aligned with the data split described in the literature for the JL1 cropland change-detection benchmark (Second Track / JL1-Second style layout).

Source

Resource	URL
JL1 Mall contest information	contest page
JL1 data / resources portal	resrepo

Data are provided by the JL1 / Jilin-1 ecosystem and the “耕地变化检测” (cropland change detection) competition. This Hub snapshot organizes released splits into a single semantic change detection (SCD) directory layout (T1/T2 + semantic labels + change mask).

Credit & citation

The train / validation labeling and split follow the convention used in remote-sensing SCD work on this benchmark. If you use this dataset in research, please cite:

@article{wangCrossDifferenceSemanticConsistency2024,
  title   = {Cross-Difference Semantic Consistency Network for Semantic Change Detection},
  author  = {Wang, Qi and Jing, Wei and Chi, Kaichen and Yuan, Yuan},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  volume  = {62},
  pages   = {1--12},
  year    = {2024},
  doi     = {10.1109/TGRS.2024.3386334}
}

Please also acknowledge the JL1 competition and data provider in-line (e.g. “JL1 CUP / JL1 Mall”) whenever you report results.

Dataset summary

Item	Value
Task	Semantic change detection (semantic maps at T1 & T2 + change mask)
Patch size	256 × 256 × 3 (uint8 PNG)
Train samples	4 050
Validation samples	1 950 (with labels)
Test samples	2 000 (bi-temporal images only, no public ground truth)

Splits are identified by full paths (train/…, val/…, test/…). The same numeric filename stem (e.g. 00001) does not denote the same geographic patch across splits.

Folder layout

.
├── train/
│   ├── metadata.csv   # Hub viewer: one row per patch → T1, T2, GT_T1, GT_T2, GT_CD columns
│   ├── T1/            # pre-change RGB
│   ├── T2/            # post-change RGB
│   ├── GT_T1/         # single-channel semantic map, T1
│   ├── GT_T2/         # single-channel semantic map, T2
│   └── GT_CD/         # binary change: 0 = no change, 255 = change
├── val/
│   ├── metadata.csv   # same column layout as train/
│   └── (same modality folders as train/)
├── test/
│   ├── metadata.csv   # Hub viewer: T1 + T2 only (no public GT)
│   ├── T1/
│   └── T2/            # hold-out evaluation; labels not distributed
├── train.txt     # one stem per line, **without** `.png`
├── val.txt       # one filename per line, **with** `.png`
└── test.txt      # same naming convention as val.txt

Resolving paths: for train, append .png to each line of train.txt under T1/ and T2/. For val and test, use each line as the filename under T1/ and T2/.

Hugging Face Dataset Viewer

Patches are multi-image samples (bi-temporal RGB + up to three label maps). If the Hub scanned the modality folders directly, T1 / T2 / GT_* would be misread as image-class subfolders and the table would not group one patch per row.

This revision fixes that by:

1. train/metadata.csv and val/metadata.csv — CSV rows with path columns t1_file_name, t2_file_name, gt_t1_file_name, gt_t2_file_name, gt_cd_file_name. Each value is relative to that split’s directory (the folder containing metadata.csv), e.g. T1/00001.png, not train/T1/00001.png.
2. test/metadata.csv — same idea but only t1_file_name and t2_file_name (no public test labels).
3. YAML configs.data_files in this card (header above) — points each split at the right metadata.csv; drop_labels: true avoids treating T1 / T2 folder names as classification labels if the Hub scans the tree. Column names use the *_file_name pattern so features resolve to Image and the viewer can render thumbnails.

Regenerate the CSVs after you change train.txt / val.txt / test.txt:

# from the SCD-CropLand-HZ repo root
python scripts/generate_jl1_second_hub_metadata.py --root datasets/JL1_second

Semantic classes (GT_T1 / GT_T2)

Index	Class
0	background
1	cropland
2	road
3	forest-grass
4	building
5	other

GT_CD: 0 = no change, 255 = change.

Change-type encoding (competition labels 0–8) maps to the above land-cover pairs (cropland ↔ road, forest, building, other, etc.); index 0 indicates no cropland-related change.

Usage

Install

pip install huggingface_hub datasets  # datasets optional, for integration

Download with the Hub

from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(
    snapshot_download(
        repo_id="BiliSakura/JL1-CUP-2024-Second-Format",
        repo_type="dataset",
    )
)
# Example paths:
# root / "train" / "T1" / "00001.png"
# root / "val.txt"  # read line-by-line for val split

You can also clone with Git LFS after installing git-lfs:

git lfs install
git clone https://huggingface.co/datasets/BiliSakura/JL1-CUP-2024-Second-Format

Limitations

• Test split: only T1 and T2 are released; there is no public GT_* for test/.
• Basenames are not comparable across train, val, and test.
• Distribution and commercial use may be restricted by JL1 Mall / competition policies—check the official pages above.

Metadata file

Machine-readable config for this revision is mirrored in the YAML block at the top of this file (dataset card header).