🌍 GTPBD-MM

A Global Terraced Parcel and Boundary Dataset with Multi-Modality

GTPBD-MM is a global multimodal remote sensing dataset for terraced parcel extraction and understanding. It extends our previous dataset GTPBD, which was released in NeurIPS 2025, by further introducing textual descriptions and Digital Elevation Model (DEM) data for each sample. The original GTPBD dataset is available at:

GTPBD-MM integrates optical imagery, text descriptions, and terrain information, making it a multimodal benchmark for studying complex terraced agricultural landscapes. The dataset provides three levels of annotations, including parcel, mask, and boundary, and supports evaluation under different settings such as Image-only, Image+Text, and Image+Text+DEM.

✨ Highlights

• Built upon GTPBD: GTPBD-MM is an extended multimodal version of our NeurIPS 2025 dataset GTPBD.
• Multi-modality: Each sample includes remote sensing imagery, textual descriptions, and DEM data.
• Fine-grained annotations: The dataset provides parcel-level, region-level, and boundary-level labels.
• Flexible experimental settings: Supports Image-only, Image+Text, and Image+Text+DEM evaluation protocols.
• Geographic metadata preserved: Each PNG image in the images directory is accompanied by an XML file containing its geographic information.

📁 Data Contents

Each sample may include:

• Optical image
• Text description
• DEM
• Parcel annotation
• Mask annotation
• Boundary annotation

In addition, every PNG image in the images/ folder is paired with a corresponding XML file that stores geographic metadata, allowing users to preserve or recover spatial reference information when needed.

🖼️ Data Format

The current public release is provided in PNG format for convenient access and usage.

Geographic Information

Although the imagery is released in PNG format, each image file is associated with an XML metadata file that records its geographic information.

Original GeoTIFF Files

If you need the original GeoTIFF (.tif) imagery for research purposes, please contact:

Zhiwei Zhang
Email: zhangzhw65@mail2.sysu.edu.cn

⚙️ Supported Research Settings

GTPBD-MM is designed to support systematic evaluation in the following settings:

• Image-only: using only optical imagery
• Image+Text: combining visual and textual information
• Image+Text+DEM: jointly exploiting imagery, language, and terrain geometry

This benchmark is intended to facilitate research on how visual appearance, textual semantics, and terrain geometry can work together to improve the understanding of terraced parcel scenes.

💡 Potential Research Topics

GTPBD-MM can support research in, but is not limited to:

• Terraced parcel extraction
• Multimodal remote sensing understanding
• Boundary-aware segmentation
• Terrain-guided vision models
• Vision-language-terrain learning
• Cross-region generalization in agricultural mapping

🔄 Updates

The associated paper is available as an arXiv preprint. We will continue to update this repository with more documentation, benchmark settings, and related resources.

📧 Contact

For questions, collaborations, or requests for the original GeoTIFF files, please contact:

Zhiwei Zhang
Email: zhangzhw65@mail2.sysu.edu.cn

📚 Citation

If you find this dataset useful in your research, please cite:

@article{zhang2026gtpbd,
  title={GTPBD-MM: A Global Terraced Parcel and Boundary Dataset with Multi-Modality},
  author={Zhang, Zhiwei and Zeng, Xingyuan and Kong, Xinkai and Zhang, Kunquan and Liang, Haoyuan and Shi, Bohan and Zheng, Juepeng and Huang, Jianxi and Lu, Yutong and Fu, Haohuan},
  journal={arXiv preprint arXiv:2604.12315},
  year={2026}
}