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
UAV hyperspectral mine screening must account for how many false alarms are reviewed before targets are found. This study compares signature sources using four detectors and simulated operator verification, finding that ACE confirmed all seven regions after nine inspections while SAM required thousands of reviews for final locations.
Problem
Method
The study compares external, fully informed in-scene, and simulated bootstrap signatures with four detectors, updating the bootstrap signature using verified detector-proposed locations.
Source: §2.4 · p. 2Results
Takeaways & Limitations
Target-discovery curves and candidate-review counts should accompany pixelwise metrics because they show how much non-target area must be inspected before each physical target is found.
Source: §4 · p. 4The bootstrap is retrospective rather than deployed, ground truth only emulates human confirmation, and the fully informed core signature is not a field-available input.
Source: §4 · p. 4
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. INTRODUCTION
The paper examines how target-signature selection affects UAV PFM-1 screening and whether operator verification can recover the benefits of an in-scene signature. It compares signatures under fixed imaging and detector conditions while measuring pixel-level performance and false-alarm burden before discovery.
1. INTRODUCTION
- HSI spectra can help separate surface materials, while UAV surveys can reduce personnel exposure in contaminated areas; recent VNIR datasets enable controlled PFM-1 studies [4] [5].The cited datasets include calibrated imagery, pixel labels, and field spectroradiometer measurements.Source: §1 · p. 1
- Scene conditions can shift target spectra, motivating signature adaptation; the experiment focuses on classical detectors requiring a target signature and background statistics because scene-specific labeled training data are assumed scarce.The cited mismatch factors include illumination, view geometry, surface condition, sensor response, reflectance retrieval, noise, and target–background mixing.Source: §1 · p. 1
- The study compares signature sources across seven physical target regions with the image, detectors, and background statistics fixed, reporting pixel-level metrics and false-alarm burden before discovery.Its operational question is whether operator verification of detector-proposed locations can recover the advantages of an in-scene target signature; the work claims neither a new detector nor a new adaptation theory.Source: §1 · p. 1
2. DATA AND METHODS
The methods evaluate PFM-1 detection in an extremely imbalanced UAV hyperspectral scene using four spectral detectors, then compare region-level discovery and spatial review across external, in-scene, and bootstrapped signatures.
2.1. UAV hyperspectral scene
- The 1705 × 3461-pixel UAV VNIR scene contains 272 bands and 248 PFM-1 pixels among 5,900,757 background pixels, a target fraction of approximately 0.0042% [4] [5].A binary mask defines the visible target support.Source: §2.1 · p. 1
2.2. Notation and target detectors
- SAM uses uncentered spectral cosine scoring, with mean-centered SAM reported separately; MF and ACE use centered vectors and regularized inverse covariance, while CEM uses standardized pixels and inverse correlation.All detector outputs are score maps where larger values indicate stronger target evidence.Source: §2.2 · p. 2
2.3. Region-level discovery evaluation
- Discovery analysis merges nine connected-component fragments into seven target regions and tracks first-positive-pixel ranks, false alarms above each region, cumulative discoveries, and spatial candidate reviews.The evaluation distinguishes pixel-level ranking from an operator inspecting non-maximum-suppressed candidate locations.Source: §2.3 · p. 2
2.4. Signature cases and bootstrap protocol
- The detector proposes the top six spatially distinct unreviewed locations; verified targets contribute central pixels to the next-round in-scene signature, while no verified target leaves the signature unchanged.The central 50% of each confirmed target region is used for signature averaging; when all seven regions are confirmed, 126 core pixels contribute to the final signature. Experiments used N = 6, ρ = 25 pixels, and η = 12 pixels; N = 6 sets a small operator-review batch before each update.Source: §2.4 · pp. 2–3
3. RESULTS
ACE leads the final bootstrap results in average precision and target discovery, while full-review bootstrapping matches the core in-scene signature after all seven regions are verified. Candidate-review effort differs sharply across detectors, and spatial maps show distinct false-alarm patterns.
3. RESULTS
- ACE has the highest AP and earliest seven-region discovery; MF and CEM retain high ROC-AUC but lower AP, while ACE gains most from in-scene refinement.The ground SVC signature gives CEM its strongest initial AP, and the full-review bootstrap matches the core in-scene case after all seven regions are verified. Fig. 2 shows precision–recall behavior under severe class imbalance.Source: §3 · p. 3, Figure 2 · p. 3, Table 1 · p. 4
- ACE found all seven regions in round 2 after 9 candidate inspections, versus 22 for CEM and 38 for MF; SAM variants needed 4558 and 2897 reviews for the last region.The discovery curves show how quickly each detector reaches the seven physical target regions as false alarms accumulate; the SAM variants’ late reviews indicate poor late-rank ordering.Source: Figure 3 · p. 3, §3 · pp. 3–4, Table 2 · p. 4
- At the common 99.5th-percentile threshold, SAM maps show broad false alarms and miss one region, while ACE has compact high scores covering all regions; support objects could be masked before review [4].MF and CEM concentrate more high scores near the target line but retain isolated false alarms; several larger false-alarm regions coincide with calibration or georeferencing objects such as panels, ground control points, or Aero-Points [4].Source: §3 · p. 4
4. DISCUSSION
HSI mine screening should be evaluated with operator-facing discovery and review measures alongside pixelwise separability, because strong ROC-AUC can coexist with false positives ahead of target regions. The retrospective bootstrap emulates human confirmation rather than deployment, and practical review requires budget-aware stopping and further testing.
4. DISCUSSION
- Target-discovery curves and candidate-review counts show how much non-target area must be inspected before each physical target is found, which ROC-AUC alone does not capture.High-ranking false positives can precede some target regions even when ROC-AUC remains strong.Source: §4 · p. 4
- The bootstrap is retrospective, not a deployed autonomous system; ground truth emulates human confirmation, and the fully informed core in-scene signature is not field-available.Verified in-scene target pixels can nevertheless reduce signature mismatch.Source: §4 · p. 4
- Practical systems should track reviewed locations and stop by an inspection budget, calibrated score threshold, or patience rule; future work should test adaptive selection, diversity constraints, and multi-detector fusion.Stopping after only the first small candidate batch may reject detectors whose first target appears deeper in the spatial ranking.Source: §4 · p. 4
5. CONCLUSION
The study finds that operational PFM-1 detection depends on signature quality and early false-alarm behavior, with ACE requiring the fewest candidate reviews to confirm all seven target locations. It recommends reporting target-discovery curves, candidate-review counts, and false-alarm-before-detection values alongside aggregate detection scores.
5. CONCLUSION
- On UAV hyperspectral imagery of PFM-1 targets, ACE confirmed all seven target locations in round 2 after nine reviewed candidates; CEM and MF required 22 and 38 reviews, while SAM variants needed thousands for final locations.Source: §5 · p. 4
- Classical detectors remain useful baselines, but their operational value depends on signature quality and early false-alarm behavior.The case study evaluated SAM, MF, ACE, and CEM with external, fully informed in-scene, and simulated human-in-the-loop signature settings.Source: §5 · p. 4
- Demining-oriented HSI studies should report target-discovery curves, candidate-review counts, and false-alarm-before-detection values because inspection burden matters alongside aggregate detection scores.Source: §5 · p. 4