FDM-YOLO: Real-time small-target UAV wildlife detection via attention-guided cross-modality fusion

Accurate monitoring of wildlife using unmanned aerial vehicles (UAVs) faces persistent challenges, particularly in distinguishing small-scale targets from complex backgrounds and effectively fusing complementary information from different sensors. To address these issues, a spatiotemporally registered UAV-based thermal infrared–RGB paired images dataset (TIR-RGB PairSet) was first constructed. Building upon this benchmark, we proposed FDM-YOLO, a domain-specific detection framework tailored for UAV bio-surveillance. To overcome the extreme scale variance of wildlife targets in drone imagery, we re-engineered the feature pyramid architecture by pruning the redundant P 5 head and introducing a high-resolution P 2 head to recover lost details of minute targets. The C3ECA module was employed to resolve the semantic misalignment between thermal signatures and RGB textures. Furthermore, we integrated SimSPPF to maximize the receptive field for global context aggregation, vital for distinguishing camouflaged animals, while significantly reducing computational FLOPs compared to standard SPPF. Experimental results demonstrate that the proposed model achieves an average of 84.71% mAP@0.5 and 44.23% mAP@0.5–0.95 on TIR-RGB PairSet after five independent trainings, significantly outperforming the YOLOv8l baseline and other recent approaches, such as YOLOv12l and YOLOv13l, while maintaining real-time processing performance. These findings highlight the potential of our framework as a reliable technical tool for UAV-based ecological monitoring, biodiversity assessment, and wildlife conservation practices. • A spatiotemporally aligned TIR–RGB wildlife dataset for UAV detection is built. • Developing FDM-YOLO with C3ECA, SimSPPF, and P2 head for dual-modality fusion. • Achieving average 84.71% mAP@0.5 and real-time detection on TIR–RGB PairSet.