Fire-YOLO26: an NMS-free dual-stream RGB–thermal network for real-time UAV forest fire detection in smoke-challenging conditions

Early and reliable forest fire detection is essential for ecological security and emergency response, yet UAV-based detectors often suffer from NMS-induced latency and modality-specific failures under smoke occlusion and thermal distractors. This study proposes Fire-YOLO26, an end-to-end NMS-free dual-stream RGB–thermal infrared (TIR) network built on YOLO26 for real-time UAV fire monitoring. A parallel RGB–TIR backbone preserves one-to-one matching inference, while a Thermal-guided Feature Alignment Module (T-FAM) derives a heat-driven spatial mask from TIR features to recalibrate RGB representations and suppress fire-like background noise. To enhance early small-fire learning, a Dual-modal Small Target Aware Labeling (Dual-STAL) strategy augments label assignment with thermal consistency. On the FLAME 2 multispectral UAV dataset, Fire-YOLO26-s achieves 94.6% mAP@0.5% and 74.2% mAP@0.5:0.95, improving mAP@0.5 by 6.5 points over the RGB-only YOLO26-s baseline, and raises heavy-smoke performance from 71.5% to 95.1% mAP@0.5. The model runs at 23 FPS on a Jetson Orin NX, indicating practical onboard deployment for all-weather early fire detection.