A lightweight and multi-scale fire detection model using YOLOv11n

Abstract

Unmanned aerial vehicles (UAVs), owing to their high efficiency and flexibility, have become vital tools for monitoring forest and agricultural fires. However, fire targets from UAV perspectives are typically small-scale, sparsely distributed, and highly variable in shape, while complex backgrounds, weather variations, and smoke interference make it challenging for existing detection models to balance detection performance and inference efficiency. Based on the YOLOv11n model, we propose Fire-YOLOv11, a lightweight fire detection architecture tailored for multi-scale flame and smoke detection. Specifically, the ShuffleNetV2 network was introduced to reconstruct the backbone for reducing parameters and computational complexity. The C3k2_MSCB module was designed to enhance the perception of flame features across different scales. The ADown module was adopted to improve feature extraction efficiency, and SELayerV2 was incorporated to strengthen the representation of salient fire-related features. Experimental results show that, compared with YOLOv11n, Fire-YOLOv11 relatively reduces model size, parameters, and GFLOPs by 54.65%, 26.60%, and 20.63%, respectively, while improving recall, mAP50, mAP50-95, and F1-score by 16.97%, 11.17%, 7.00%, and 7.68%, respectively. Further multi-seed experiments and validation on a public UAV fire dataset demonstrate that Fire-YOLOv11 maintains stable detection performance while substantially reducing model complexity. These results suggest that the proposed model achieves a reasonable trade-off between detection performance and deployment efficiency, highlighting its potential for real-time fire monitoring on resource-constrained UAV platforms.