Metallic landmine detection and identification using machine learning classification of UAV-based magnetometer data

A total of 202 magnetic anomaly signals were extracted from four low-altitude UAV flights over a single test site containing 43 buried metallic landmines and debris. Percentile distance, defined as the distance at which an anomaly decreases to a specified percentage of its maximum, was introduced to characterize signal attenuation.

Five machine-learning models were evaluated using maximum anomalies and percentile-distance features. Repeated stratified cross-validation showed that selected percentile distances improved macro-F1 scores for all models compared with the full percentile range. SVM achieved the highest macro-F1 score (0.9586 ± 0.0372), and hold-out testing showed no confusion between landmines and debris.