README.md

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

# Forest-Change

## Overview
Forest-Change is the first benchmark dataset specifically designed for joint forest change detection and captioning in remote sensing imagery. It provides bi-temporal satellite images, pixel-level deforestation masks, and multi-granularity semantic captions describing forest cover changes in tropical and subtropical regions.

## Dataset Details
- **Total Examples**: 334 annotated bi-temporal image pairs
- **Spatial Resolution**: ~30m/pixel (medium resolution)
- **Original Image Size**: 480×480 pixels (cropped from larger scenes)
- **Processed Image Size**: 256×256 pixels (resized for model training)
- **Temporal Resolution**: 1 year between image pairs
- **Geographic Focus**: Tropical and subtropical deforestation fronts

## Dataset Splits
- **Training**: 270 examples (~80%)
- **Validation**: 31 examples (~10%)
- **Test**: 33 examples (~10%)

## 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=deforestation)
- **Captions**: Five captions describing the forest change event with varied granularity

## Data Sources
- **Imagery Source**: Google Earth Engine (GEE)
- **Base Dataset**: Derived from Hewarathna et al. (2024) forest ecosystem change detection dataset
- **Validation**: Forest cover changes verified through Global Forest Watch (GFW) platform
- **Geographic Selection**: Based on WWF 2015 Deforestation Fronts report

## Caption Generation
Captions are generated through a hybrid two-stage approach:
1. **Human Annotation**: One caption per example manually created by domain annotators describing observed changes
2. **Rule-Based Generation**: Four additional captions automatically generated based on quantitative mask properties:
   - Percentage of newly deforested area (binned into descriptive severity levels)
   - Size and number of individual change patches
   - Spatial distribution patterns of deforestation
   - Variation in patch sizes

This approach ensures both semantic richness from human expertise and consistent structural variation across captions.

## Key Characteristics
- **Change Coverage**: 
  - Mean: <5% deforestation per image
  - Maximum: 40% deforestation
  - Distribution: Heavily skewed toward lower deforestation percentages
- **Caption Length**: Bimodal distribution with both concise and detailed descriptions
- **Change Patterns**: Diverse deforestation manifestations including:
  - Scattered small patches across forest areas
  - Concentrated clearing zones
  - Edge-of-clearing expansion patterns
  - Highly variable patch sizes and configurations
- **Caption Content**: Descriptions emphasize:
  - Degree/severity of forest loss
  - Spatial location within the image
  - Patch characteristics (size, number, distribution)

## 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 normalization using dataset-specific mean and standard deviation statistics
- No atmospheric correction applied
- No cloud masking applied (some samples contain partial cloud occlusion)

## Use Cases
- Forest change detection and monitoring
- Deforestation segmentation in natural environments
- Change captioning for ecological applications
- Multi-task learning for remote sensing
- Benchmarking vision-language models on forest imagery
- Training interactive forest analysis systems
- Developing automated forest monitoring workflows

## Evaluation Metrics
Due to severe class imbalance (most pixels are no-change), evaluation requires:
- **Per-class IoU**: Separate metrics for change and no-change classes
- **Mean IoU (mIoU)**: Average of both class IoUs
- **Caption metrics**: BLEU-n (n=1,2,3,4), METEOR, ROUGE-L, CIDEr-D
- **Note**: Overall accuracy is not recommended due to class imbalance

## Limitations
- **Dataset Scale**: Limited to 334 examples, restricting model generalization
- **Scene Diversity**: Limited number of unique geographic sites due to cropping and augmentation strategy
- **Class Imbalance**: Severe imbalance with most pixels representing no-change, challenging for detection models
- **Caption Quality**: Majority of captions are rule-based, limiting linguistic variation and naturalness
- **Geographic Grounding**: Limited incorporation of geographic features and contextual information in captions
- **Spatial Resolution**: Medium resolution (~30m/pixel) limits detection of very small-scale changes
- **Temporal Coverage**: Fixed 1-year intervals between image pairs
- **Atmospheric Effects**: Some samples affected by partial cloud occlusion
- **Edge Boundaries**: Fuzzy boundaries at deforestation patch edges complicate precise segmentation

## Citation
If you use this dataset, please cite:
```bibtex
@article{brock2024forestchat,
  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{hewarathna2024change,
  title={Change detection for forest ecosystems using remote sensing images with siamese attention u-net},
  author={Hewarathna, AI and Hamlin, L and Charles, J and Vigneshwaran, P and George, R and Thuseethan, S and Wimalasooriya, C and Shanmugam, B},
  journal={Technologies},
  volume={12},
  number={9},
  pages={160},
  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: james.brock@bristol.ac.uk
- School of Computer Science, University of Bristol