JimmyBrocko/JL1-CD-Trees

---
license: mit
task_categories:
- image-segmentation
- image-to-image
language:
- en
size_categories:
- 1K<n<10K
---

# JL1-CD-Trees

## Overview
JL1-CD-Trees is a curated subset of the JL1-CD dataset filtered for tree and woodland cover changes. The source dataset covers a diverse range of geographic regions and land cover types across China, including human-induced and natural surface changes. It supports change detection tasks only — no change captions are included.

## Dataset Details
- **Source**: Filtered subset of the JL1-CD dataset (Liu et al., 2025)
- **Geographic Coverage**: Multiple provinces across China (Shandong, Ningxia, Anhui, Hebei, Hunan and others)
- **Temporal Range**: Early 2022 to end of 2023

## Dataset Splits
- **Training**: 244
- **Validation**: 81
- **Test**: 83

## Data Format
Each example contains:
- **Image A**: Pre-change RGB satellite image
- **Image B**: Post-change RGB satellite image
- **Change Mask**: Binary segmentation mask (0=no change, 1=change)

## Filtering Criteria
Examples are selected from JL1-CD based on scene content, retaining image pairs containing visible tree or forest cover changes.

## Key Characteristics
- **Change Coverage**:
- Mean: 5.04% per image
- Maximum: 48.79%
- **Annotation Focus**: Binary pixel-level change masks
- **Caption Support**: None — change detection only
- **Object Geometry**: Mixed patterns including urban infrastructure, grassland, and tree cover boundaries

## Preprocessing
- All images resized to 256×256 pixels for consistency
- Change masks binarized (0=no change, 1=change)
- Bi-temporal image pairs pre-aligned
- Per-channel normalisation using dataset-specific mean and standard deviation statistics

## Use Cases
- Remote sensing change detection across diverse geographic regions and land cover types
- Cross-domain transfer learning from forest to mixed land cover scenes
- Benchmarking model generalisation on high-resolution imagery
- Training and evaluating interactive remote sensing agents

## Evaluation Metrics
- **Per-class IoU**: Separate metrics for change and no-change classes
- **Mean IoU (mIoU)**: Average of both class IoUs
- **Note**: Overall accuracy is not recommended due to class imbalance

## Limitations
- **No captions**: Change detection only — captioning tasks are not supported
- **Seasonal and atmospheric variation**: Imagery contains variable atmospheric and seasonal conditions which may affect model performance
- **Fixed image size**: 256×256 pixels
- **Resolution**: High-resolution imagery (0.5-0.75m/pixel) may not generalise to medium-resolution datasets

## Citation
If you use this dataset, please cite:
```bibtex
@article{brock2026forest,
title={Forest-Chat: Adapting Vision-Language Agents for Interactive Forest Change Analysis},
author={Brock, James and Zhang, Ce and Anantrasirichai, Nantheera},
journal={arXiv preprint arXiv:2601.14637},
year={2026}
}

@article{liu2025jl1,
title={JL1-CD: A new benchmark for remote sensing change detection and a robust multi-teacher knowledge distillation framework},
author={Liu, Ziyuan and Zhu, Ruifei and Gao, Long and Zhou, Yuanxiu and Ma, Jingyu and Gu, Yuantao},
journal={arXiv preprint arXiv:2502.13407},
year={2025}
}
```

Paper information available at: https://huggingface.co/papers/2601.14637.

## License
MIT License - Academic re-use purpose only

## Contact
For questions or issues regarding this dataset, please contact:
- James Brock: [e-mail]
- School of Computer Science, University of Bristol