Gtpbd Eval

gtpbd-eval

GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset — Zhiwei Zhang et al. (2025) (arXiv:2507.14697, 2025)

What this evaluates

Evaluates fine-grained agricultural parcel delineation, boundary detection, and cross-domain generalization on high-resolution remote sensing imagery of terraced terrain. It benchmarks semantic segmentation, edge extraction, and parcel extraction models across multiple geographic domains.

Datasets

• GTPBD — total 47537; splits: train (-1), val (-1), test (-1); repo https://github.com/Z-ZW-WXQ/GTPBD

Metrics

• IoU (primary) — range: [0, 1]
  • Intersection over Union: ratio of the intersection area between predicted and ground truth masks to their union area.
• F1-score — range: [0, 1]
  • Harmonic mean of Precision and Recall: 2 * (Prec * Rec) / (Prec + Rec).
• ODS — range: [0, 1]
  • Optimal Dataset Scale F1-score: F1 computed using a single threshold optimized over the entire dataset.
• OIS — range: [0, 1]
  • Optimal Image Scale F1-score: F1 computed using the best threshold per image, then averaged across the dataset.
• AP — range: [0, 1]
  • Average Precision: area under the precision-recall curve across multiple thresholds.
• GOC — range: percent
  • Global Over-Classification Error: percentage of parcels incorrectly split or merged, measuring over-segmentation at the object level.
• GUC — range: percent
  • Global Under-Classification Error: percentage of parcels missed or merged, measuring under-segmentation at the object level.
• GTC — range: percent
  • Global Total Classification Error: sum of GOC and GUC, representing total object-level classification error.

Input / output format

Input: 512×512 cropped patches of high-resolution remote sensing imagery covering terraced agricultural terrain.

Output: Pixel-level segmentation masks, edge probability maps (binarized), or object-level parcel masks/polygons.

Scoring recipe

def compute_iou_f1(pred_mask, gt_mask, thr=0.5):
    pred_bin = (pred_mask >= thr).astype(int)
    inter = np.logical_and(pred_bin, gt_mask).sum()
    union = np.logical_or(pred_bin, gt_mask).sum()
    iou = inter / union if union > 0 else 0.0
    tp = inter
    fp = np.logical_and(pred_bin, ~gt_mask).sum()
    fn = np.logical_and(~pred_bin, gt_mask).sum()
    prec = tp / (tp + fp) if (tp + fp) > 0 else 0
    rec = tp / (tp + fn) if (tp + fn) > 0 else 0
    f1 = 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0
    return iou, f1

def compute_ods(pred_maps, gt_edges):
    # Optimize single threshold over all images to maximize F1
    pass

def compute_ois(pred_maps, gt_edges):
    # Optimize threshold per image, then average F1
    pass

Common pitfalls

• Edge ground truth is synthetically generated via morphological erosion/dilation (1–5 px widths) rather than manually traced boundaries.
• Object-level metrics (GOC/GUC/GTC) measure parcel classification errors (over/under-segmentation) rather than boundary distance errors.
• UDA experiments use specific geographic domain pairs (S, N, G) representing distinct terraced zones, not standard synthetic-to-real splits.

Evidence (verbatim from paper)

To evaluate segmentation accuracy on the GTPBD dataset, we adopt five standard pixel-level metrics commonly used in semantic segmentation and unsupervised domain adaptation (UDA) tasks, including Precision (Prec.), Recall (Rec.), Intersection over Union (IoU), Overall Accuracy (OA) and F1-score;

Citation

@misc{zhang2025gtpbd,
  title={GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset},
  author={Zhiwei Zhang et al. (2025)},
  year={2025},
  note={arXiv:2507.14697}
}

• arXiv: 2507.14697