Pfm1 Landmine Detection Eval

pfm1-landmine-detection-eval

Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery — Lekhak et al. (2026) (arXiv:2602.10434, 2026)

What this evaluates

Evaluates the ability of statistical and learning-based detectors to identify sparse PFM-1 landmines in UAV-captured hyperspectral imagery, emphasizing performance under severe class imbalance and varying background clutter.

Datasets

• UAV Hyperspectral Imagery (PFM-1 Landmine Scene) — total ?; splits: Full Region (-1), PFM-1 Region (-1), Test Region (-1)

Metrics

• AUC — range: [0, 1]
  • Area under the Receiver Operating Characteristic curve, measuring the trade-off between true positive rate and false positive rate across all classification thresholds.
• AP (primary) — range: [0, 1]
  • Average Precision, computed as the area under the Precision-Recall curve. It summarizes detector performance under severe class imbalance by weighting precision at each recall level.

Input / output format

Input: VNIR hyperspectral image cubes (pixel-wise spectral signatures) from UAV flights over landmine test sites.

Output: Per-pixel detection scores (continuous values) used to generate binary detection masks at varying thresholds for ROC and PR curve computation.

Scoring recipe

def compute_metrics(scores, labels):
    fpr, tpr, _ = roc_curve(labels, scores)
    auc = auc(fpr, tpr)
    precision, recall, _ = precision_recall_curve(labels, scores)
    ap = auc(recall, precision)
    return {'AUC': auc, 'AP': ap}

Common pitfalls

• Relying solely on ROC-AUC, which remains high even when precision is poor due to extreme target sparsity and background clutter.
• Ignoring scene composition effects; AP drops significantly in the Full Region due to false alarms from non-target objects (e.g., aeropoints), masking true detector capability.
• Evaluating on cropped regions only without an independent Test Region, leading to overoptimistic results that do not generalize to unseen scenes.

Evidence (verbatim from paper)

Precision–recall analysis, presented in Fig. [4] provides complementary insight under severe class imbalance and reveals detector behavior not apparent from ROC curves alone. For the Full Region, all methods exhibit low Average Precision (AP), reflecting extreme target sparsity and high background diversity. These results highlight that aggregate AUC alone is insufficient for evaluating rare-target detection performance and that low-FPR behavior is a key discriminator for landmine detection applications.

Citation

@misc{lekhak2026pfm1detection,
  title={Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery},
  author={Lekhak et al. (2026)},
  year={2026},
  note={arXiv:2602.10434}
}

• arXiv: 2602.10434