PFM-1 Landmine Detection in Vegetation Using Thermal Imaging with Limited Training Data

Landmine detection, especially of PFM-1 “butterfly’’ mines, remains a critical challenge in post-conflict environments due to their small size and plastic construction. Modern demining operations increasingly incorporate advanced sensing modalities, including multi-spectral imaging. Long-Wave Infrared (LWIR), in particular, offers potential for detecting PFM-1 mines under certain environmental conditions. However, the limited availability of annotated thermal data restricts the generalization capabilities of deep learning approaches. To address these limitations, we propose a multi-stage, feature-based detection algorithm tailored for PFM-1 mines in LWIR thermal imagery. The method is trained on Track 1 and Track 2 of the MineInsight dataset, using F1 and F2 score-weighted loss functions to prioritize recall and reduce false negatives. Across ten independent runs, the proposed method exhibits comparable performance and complementary robustness relative to YOLOv8, especially under conditions of limited training data or partial occlusion.