LEVIR-MCI-Trees

Overview

LEVIR-MCI-Trees is a curated subset of the LEVIR-MCI dataset specifically focused on tree cover changes in urban and peri-urban environments. This dataset supports joint change detection and captioning tasks for remote sensing imagery, containing bi-temporal image pairs with pixel-level change masks and semantic descriptions.

Dataset Details

• Source: Filtered subset of LEVIR-MCI dataset (Liu et al., 2024)
• Total Examples: 2,305 image pairs
• Spatial Resolution: 0.5m/pixel
• Image Size: 256×256 pixels
• Temporal Range: 5-15 years between image pairs
• Geographic Focus: Urban and peri-urban areas with tree cover changes

Dataset Splits

• Training: 1,518 examples (66%)
• Validation: 374 examples (16%)
• Test: 413 examples (18%)

Data Format

Each example contains:

• Bi-temporal image pairs: Two RGB images (Image A and Image B) captured at different time points
• Change mask: Binary/multi-class segmentation mask highlighting changes to roads and buildings
• Captions: Five human-annotated captions describing the observed changes from varying perspectives

Filtering Criteria

Examples are selected from LEVIR-MCI based on caption content. An image pair is included if at least one of its five captions contains tree-related keywords: 'tree', 'trees', 'wood', 'woods', 'woodland', 'wooded', 'forest', 'forests', 'jungle', or 'jungles'.

Key Characteristics

• Change Coverage: Mean 15.28%, maximum 72.79% of image area
• Annotation Focus: Pixel-level annotations for roads and buildings (not trees directly)
• Caption Style: Concise descriptions with diverse vocabulary and varied perspectives
• Object Geometry: Regular geometric patterns characteristic of urban infrastructure and managed landscapes
• Image Quality: High-resolution imagery suitable for fine-grained analysis

Use Cases

• Remote sensing change detection in urban environments
• Change captioning and description generation
• Multi-task learning for vision-language models
• Benchmarking model performance on high-resolution imagery
• Urban forest monitoring and tree cover analysis
• Training and evaluating interactive remote sensing agents

Limitations

• Change masks only annotate roads and buildings, not tree cover changes directly
• Limited to high-resolution imagery (0.5m/pixel)
• Fixed image size of 256×256 pixels
• Urban-focused context may not represent natural forest environments
• Variable temporal spans between image pairs (5-15 years)

Citation

If you use this dataset, please cite:

title={Forest-Chat: Adapting Vision-Language Agents for Interactive Forest Change Analysis},
author={Brock, James and Zhang, Ce and Anantrasirichai, Nantheera},
journal={Ecological Informatics},
year={2024}
}

@article{liu2024changeagent,
title={Change-Agent: Towards Interactive Comprehensive Remote Sensing Change Interpretation and Analysis},
author={Liu, Chenyang and Chen, Keyan and Zhang, Haotian and Qi, Zipeng and Zou, Zhengxia and Shi, Zhenwei},
journal={IEEE Transactions on Geoscience and Remote Sensing},
year={2024}
}

Paper information available at: https://huggingface.co/papers/2601.04497 and 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: [email protected]
• School of Computer Science, University of Bristol