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

In a UAV hyperspectral scene where PFM-1 mines are 0.0042% of pixels (248 of 5.9M), every reasonable detector looks great by ROC-AUC (>0.99 for MF/ACE/CEM). But the number of false candidates an operator must physically walk up to and inspect before all seven mines are found ranges from 9 to 4,558 depending on which detector you picked. Aggregate scores hide nearly all of the operational difference.

We compare SAM (centered and uncentered), MF, ACE, and CEM under three target-signature sources: a ground-measured SVC spectrum, a fully informed in-scene core-pixel spectrum, and a simulated human-in-the-loop bootstrap that starts from the field spectrum and refines it only after an operator confirms detector-proposed locations (6 candidates per round, NMS radius ≈0.32 m on the ground, ≈0.16 m inspection radius).

ACE confirms all 7 target regions in 2 rounds / 9 inspections. CEM needs 22, MF 38. The SAM variants find several targets early but need thousands of reviews for the last one — poor late-rank ordering.
Bootstrapping does recover the fully informed in-scene case: ACE's AP goes 0.149 → 0.314. A handful of operator-verified pixels buys most of the benefit of a signature you couldn't have had in the field.
Several of the largest false-alarm clusters turn out to be the calibration panels and ground control points we deployed ourselves — maskable in an operational workflow.

Caveats up front: retrospective, one scene, seven targets. Ground truth stands in for the operator, and the in-scene signature is a fully informed reference rather than a field-available input.

Code: https://github.com/SagarLekhak/IEEE_WHISPERS_2026_UAV_HSI_PFM1
Dataset: arXiv:2510.02700 · Prior detector benchmark: arXiv:2602.10434

Open question for anyone working on rare-target detection: is there a better single summary of inspection burden than false-alarms-before-detection? We want something comparable across scenes with different target counts, and we don't think AP is it.