Smoldering hotspots that persist after wildfire suppression represent one of the most critical sources of reignition risk. This study evaluates the precision and false positive rate of YOLO (You Only Look Once) object detection models — specifically YOLOv5, YOLOv7, YOLOv8, and YOLOv11 — in detecting post-fire smoldering hotspots from infrared imagery acquired by thermal cameras integrated into unmanned aerial vehicles (UAVs/drones). The primary motivation stems from the inadequacy of traditional ground-based patrol methods in scanning vast burned areas and the difficulty of distinguishing thermal anomalies with the human eye. The study utilizes the publicly available FLAME dataset alongside original thermal datasets collected through controlled field experiments. Model performance was evaluated using precision, recall, mean average precision (mAP@0.50 and mAP@0.50:0.95), and false positive rate (FPR) metrics. Findings reveal that the YOLOv8m model achieved the highest detection accuracy with a 91.4% mAP@0.50 value, while maintaining a false positive rate of 6.8%. The YOLOv5s model achieved 86.2% mAP@0.50 with lower computational cost, though its false positive rate increased to 11.3% under nighttime conditions. The study discusses the role of attention mechanisms (CBAM, EMA) and multi-scale feature extraction strategies in smoldering hotspot detection and provides recommendations for multi-modal sensor fusion in future research.
Kaan Alper (Fri,) studied this question.