Human-in-the-Loop Signature Bootstrapping for UAV Hyperspectral PFM-1 Mine Detection

Detection Without Drowning in False Alarms

Hyperspectral imaging (HSI)—capturing images across dozens of light wavelengths—can identify mines by their spectral "fingerprint," but only if operators don't exhaust themselves reviewing thousands of false positives. This paper shows that a human-in-the-loop approach, where operators iteratively verify a few promising candidates to refine the detection model, can match the performance of having perfect ground-truth data—but the computational method matters enormously. The adaptive coherence estimator (ACE) algorithm confirmed all seven mines in just nine candidate reviews across two rounds, while spectral angle mapper (SAM) required thousands of reviews to find the same targets.

The Setup and Methods

The authors flew a UAV over terrain where PFM-1 anti-personnel mines were present and collected visible and near-infrared (VNIR) hyperspectral data. They tested four spectral-matching algorithms:

• SAM (spectral angle mapper): measures angular distance between pixel spectra and a reference signature
• MF (matched filter): treats the signature like a template to correlate against the image
• ACE (adaptive coherence estimator): a probabilistic detector robust to background variability
• CEM (constrained energy minimization): suppresses energy from non-target spectra

For each algorithm, they evaluated three signature sources:

1. A ground-measured signature from a spectrometer in a lab (realistic but not from the actual scene)
2. A fully informed in-scene signature (the oracle: pixel-perfect reference extracted from the image)
3. A bootstrapped signature from iterative human inspection (simulated, where an operator confirms or rejects the top candidate detections, then the algorithm retrains)

Why This Matters

Detection quality is usually summarized by ROC curves and average precision metrics, but those hide a critical operational reality: how many pixels must a human review to find all targets? The authors therefore tracked target-discovery curves and actual candidate-review counts—metrics that directly reflect field workload.

The bootstrapping simulation revealed stark differences. ACE, which naturally handles spectral variability, converged quickly: after humans verified just nine candidates across two inspection rounds, the algorithm's performance matched the oracle case with perfect information. SAM and its variants, conversely, generated thousands of false candidates even after bootstrapping, making them impractical for real mine-clearance operations despite similar ROC scores on paper.

This finding is operationally crucial. A detector with a slightly lower ROC area but far fewer false alarms saves lives—fewer mines go undetected because operators fatigue, and safer areas are cleared faster.

The work also provides reproducible code, enabling other teams to benchmark detectors on the same UAV HSI dataset and advance mine-detection research beyond synthetic or lab settings.