UAV hyperspectral mine detection: ACE finds targets 200x faster than SAM
New human-in-the-loop bootstrap cuts false-alarm inspections from thousands to nine.
A new study from Rochester Institute of Technology explores how human-in-the-loop bootstrapping can dramatically speed up landmine detection from UAV hyperspectral imagery. The paper, led by Sagar Lekhak, compares four classic detection algorithms—spectral angle mapper (SAM), matched filter (MF), adaptive coherence estimator (ACE), and constrained energy minimization (CEM)—on the task of identifying PFM-1 mines using visible and near-infrared (VNIR) hyperspectral data.
The key innovation is a simulated human-in-the-loop signature bootstrap, where an analyst iteratively verifies or rejects algorithm detections to refine the target signature. The results show a massive performance gap: ACE confirms all seven target regions in just two rounds with only nine candidate inspections, while SAM variants require thousands of reviews to locate the final targets. The bootstrap eventually matches the performance of a fully informed in-scene signature, but the effort required varies enormously depending on the algorithm chosen.
- ACE detector needed only 9 candidate inspections across 2 rounds to confirm all 7 PFM-1 target regions.
- SAM-based detection variants required thousands of candidate reviews to find the same targets.
- Full human-in-the-loop bootstrapping achieves accuracy equal to a perfect in-scene signature after verification.
Why It Matters
Could reduce manual inspection effort from thousands to single digits for mine clearance operations.