Source-linked AI summary
Human-in-the-Loop Signature Bootstrapping for UAV Hyperspectral PFM-1 Mine Detection
Sagar Lekhak, Prasanna Reddy Pulakurthi, Emmett J. Ientilucci
TL;DR
Operational PFM-1 mine screening requires understanding how signature selection and operator review affect target discovery, not just pixel-level separability. This paper compares detector and signature configurations with a simulated human-in-the-loop bootstrap, finding that ACE confirms all seven target locations in round 2 after nine reviewed candidates, while SAM variants require thousands of reviews for their final locations.
Problem
The paper asks how target-signature selection affects UAV PFM-1 detection and whether operator verification can recover the advantages of an in-scene signature.
Method
The study compares SAM, MF, ACE, and CEM using external, fully informed in-scene, and simulated human-in-the-loop bootstrapped signatures on seven target regions.
Results
ACE confirms all seven target locations in round 2 after nine reviewed candidates, whereas SAM variants require thousands of reviews for their final target locations.
Takeaways & Limitations
HSI mine-screening evaluations should report target-discovery curves and candidate-review counts alongside pixelwise metrics because inspection burden matters operationally.
Takeaways & Limitations
The bootstrap experiment is retrospective, uses ground truth to emulate human confirmation, and treats the core in-scene signature as a field-unavailable reference.
Abstract
from arXiv · showhide
Hyperspectral imaging (HSI) is useful for material discrimination, but operational mine screening also depends on how many false alarms must be inspected before targets are found. This paper studies PFM-1 landmine detection in unmanned aerial vehicle (UAV) visible and near-infrared (VNIR) HSI using spectral angle mapper (SAM), matched filter (MF), adaptive coherence estimator (ACE), and constrained energy minimization (CEM). We compare a ground-measured SVC signature, a fully informed in-scene core-pixel signature, and a simulated human-in-the-loop signature bootstrap. Besides receiver operating characteristic area under the curve and average precision, we report target-discovery curves and spatial candidate-review counts. Full-review bootstrapping reaches the fully informed in-scene signature case after all seven target regions are verified, but the required inspection effort varies strongly: ACE confirms all regions in two rounds and nine candidate inspections, whereas the SAM variants need thousands of candidate reviews for their final target locations. Code is available at https://github.com/SagarLekhak/IEEE_WHISPERS_2026_UAV_HSI_PFM1.
1 Rochester Institute of Technology, Rochester, NY 14623, USA
This section identifies the paper’s core topics: hyperspectral imaging target detection with UAVs, PFM-1 landmine detection, and several detection algorithms with human-in-the-loop processing.
- The study addresses HSI target detection using UAV platforms.
- Its application is PFM-1 landmine detection.
- The listed algorithms are SAM, MF, ACE, and CEM, alongside human-in-the-loop analysis.
1. INTRODUCTION
The introduction motivates UAV VNIR hyperspectral imaging for stand-off PFM-1 screening and frames signature selection as an operational determinant of detector performance. The paper evaluates whether operator verification of proposed locations can recover benefits associated with an in-scene target signature, rather than proposing a new detector or adaptation theory.
- Motivation: HSI supports stand-off landmine detection by using calibrated wavelength-resolved spectra to separate surface materials.UAVs make close-range surveys more practical while reducing personnel exposure in contaminated areas.
- Motivation: Recent UAV VNIR datasets enable controlled PFM-1 studies using calibrated imagery, pixel labels, and field spectroradiometer measurements.
- Signature selection: SAM, MF, ACE, and CEM are widely used target detectors, but their performance depends on how well the available spectrum matches the target as imaged in a specific scene.Apparent signatures can shift with illumination, view geometry, surface condition, sensor response, reflectance retrieval, noise, and target–background mixing.
- Research question: The paper asks how target-signature selection affects UAV-based PFM-1 screening and whether operator verification of detector-proposed locations can recover the advantages of an in-scene target signature.The experiment is designed to emulate key aspects of a real-world screening setting with little to no scene-specific labeled training.
2. DATA AND METHODS
The study evaluates UAV VNIR hyperspectral PFM-1 detection in a severely imbalanced scene using multiple detector formulations, target signatures, and human-in-the-loop bootstrapping. Evaluation combines pixel-level metrics with target-discovery and spatial candidate-review analyses aligned with demining operations.
- Dataset: The analyzed image contains 1705 × 3461 pixels and 272 retained bands spanning approximately 400–1000 nm, with 248 positive pixels against 5,900,757 background pixels.The target fraction is approximately 0.0042%, making the scene strongly imbalanced.
- Evaluation: The evaluation supplements ROC-AUC and precision–recall measures with target-region discovery and spatial candidate-review analyses because operational demining concerns physical locations and false-alarm inspection effort.For each region, the method records the first target-pixel rank and the number of non-target pixels ranked above it, then plots regions discovered as false alarms accumulate.
- Signatures: Three signature settings were evaluated: an external SVC spectroradiometer signature, a fully informed in-scene signature averaged from central pixels across all seven target regions, and a human-in-the-loop bootstrap.The in-scene signature is a reference rather than a field-available input or strict mathematical upper bound.
- Signature bootstrap: The bootstrap initializes with the external SVC signature, proposes the top N spatially distinct unreviewed locations, and updates the signature using central pixels from verified target regions.If no target is verified, the current signature is retained; confirmed signatures use the central 50% of each target region, and all seven regions contribute 126 core pixels.
- Implementation: The experiments use N = 6, ρ = 25 pixels, and η = 12 pixels, corresponding to about 0.32 m candidate separation and a 0.16 m inspection radius.All detectors run over the full image; min-max normalization is used only for visualization and does not change rank-based statistics.
3. RESULTS
Results show that ACE provides the strongest precision and fastest target discovery, while full-review bootstrapping matches the fully informed in-scene signature after all seven regions are verified. Spatial review effort and map compactness also favor ACE over the other detectors.
- Pixel-level results: Only 248 of 5,901,005 pixels are targets, making average precision a clearer comparison than ROC-AUC under severe class imbalance.ROC-AUC is high for MF, ACE, and CEM across signature cases, but AP better distinguishes performance when background pixels dominate.
- Pixel-level results: The ground SVC signature gives CEM the strongest initial AP, while ACE shows the largest AP gain from in-scene refinement.Full-review bootstrapping matches the core in-scene case because its final signature is built after all seven target regions are verified.
- Target discovery: ACE has the highest AP and reaches all seven target regions earliest in the discovery curve.The precision-recall and cumulative-discovery views respectively show pixel-level ranking behavior and the accumulation of false alarms before physical targets are found.
- Spatial review effort: 2 rounds and 9 inspections are required for ACE to reach all seven regions, compared with 22 inspections for CEM and 38 for MF.The candidate-review bootstrap used six new spatial candidate sites per detector per round.
- Spatial review effort: 4,558 reviews for one SAM variant and 2,897 for the other are required to reach their final regions, despite several targets appearing early.These totals indicate poor late-rank ordering; at the common 99.5th-percentile threshold, SAM maps also show broad false alarms and miss one target region.
- Spatial map patterns: At the common 99.5th-percentile threshold, ACE produces the most compact high-score pattern, represents all target regions, and shows fewer high-score background structures.MF and CEM concentrate more high scores near the target line but retain isolated false alarms.
4. DISCUSSION
The discussion argues that UAV HSI mine screening should be assessed with operator-facing inspection metrics alongside pixelwise separability. It also frames the bootstrap as a retrospective proxy and recommends budgeted stopping rules for practical systems.
- Operational evaluation: Operator-facing target-discovery curves and candidate-review counts show how much non-target area must be inspected before each physical target location is found.ROC-AUC can remain strong even when high-ranking false positives precede some target regions.
- Bootstrap limitations: The bootstrap experiment is retrospective, with ground truth emulating human confirmation and the core in-scene signature serving as a fully informed reference rather than a field input.Verified in-scene target pixels can reduce signature mismatch, but early stopping after a small candidate batch may reject detectors whose first target ranks deeper.
- Operational safeguards: Practical systems should track reviewed locations and stop using an inspection budget, calibrated score threshold, or patience rule.These controls address the risk of stopping before deeper-ranked target regions are found.
5. CONCLUSION
The case study evaluated four classical detectors across external, fully informed in-scene, and simulated human-in-the-loop signature settings for UAV hyperspectral PFM-1 detection. ACE was the most efficient, confirming all seven target locations in round 2 after nine reviewed candidates, while CEM and MF required more reviews.
- Conclusion: SAM, MF, ACE, and CEM were evaluated on UAV hyperspectral imagery of PFM-1 targets under three signature settings.The settings were external, fully informed in-scene, and simulated human-in-the-loop signatures.
- Conclusion: Detector operational value depended on signature quality and early false-alarm behavior, despite classical detectors remaining useful baselines.
- Conclusion: 9 reviewed candidates enabled ACE to confirm all seven target locations in round 2, compared with 22 reviews for CEM and 38 for MF.