GTPBD (Global Terraced Parcel and Boundary Dataset) is the first fine-grained benchmark dataset for agricultural parcel extraction in global terraced regions. Unlike previous datasets focused on regular, flat farmlands, GTPBD specifically addresses complex terrain types, providing rich annotations and multi-task evaluation protocols for semantic segmentation, edge detection, parcel extraction, and domain adaptation.
-
Global Scope: 47,537 high-resolution images from China, Vietnam, Tunisia, Ethiopia, Mexico, Peru, and more.
-
Terraced Focus: Over 200,000 annotated terraced parcels, including mountainous, shadow-occluded, and irregular field structures.
-
Multi-level Labels:
- Parcel Mask Labels
- Pixel-level Boundary Labels
- Instance-level Parcel Labels
-
Multi-Task Support:
- Semantic Segmentation
- Edge Detection
- Agricultural Parcel Extraction
- Unsupervised Domain Adaptation (UDA)
-
Geographic Coverage: 7 major zones in China + global transcontinental regions
-
Area Covered: 885 km² of annotated terraces
-
Label Diversity:
- 3-level labels (mask, boundary, parcel)
- Complex topologies (shared vs. non-shared field ridges)
- Models: U-Net, DeepLabV3, PSPNet, SegFormer, Mask2Former, etc.
- Metrics: IoU, Pixel Accuracy, F1-score, Recall, Precision
- Models: UEAD, MuGE, PiDiNet, REAUNet-Sober
- Metrics: ODS, OIS, AP
-
Models: HBGNet, SEANet, REAUNet
-
Metrics: OA, IoU, F1, plus object-level:
- GOC (Over-Classification Error)
- GUC (Under-Classification Error)
- GTC (Total Classification Error)
- Domains: South (S), North (N), Global (G)
- Transfers: S → N, G → S, etc.
- Methods: Source Only, FDA, DAFormer, HRDA, PiPa
Hosted on Hugging Face: GTPBD on Hugging Face
git clone https://huggingface.co/datasets/wxqzzw/GTD
Croissant metadata is available for NeurIPS submission validation.
- Maintainer: Zhiwei Zhang (HuggingFace Profile)
- License: cc-by-nc-4.0
We thank all 50+ annotators and collaborators for their contributions to this large-scale terraced dataset effort.
For questions, feel free to raise an issue or contact the maintainers via GitHub.