Addressing the limitations of inadequate model lightweighting and suboptimal detection accuracy in forest fire detection systems, a refined forest fire detection approach based on an improved YOLO11 architecture is proposed. Based on the YOLO11 network architecture, the backbone network is modified by integrating the ShuffleNetV1 module to achieve efficient and lightweight model deployment. Additionally, the incorporation of the SPD-Conv convolutional module not only expands the receptive field to strengthen the aggregation of semantic features for large-scale flame targets but also precisely preserves fine-grained edge and texture information of small-scale smoke targets. The experimental results show that the improved model achieves a real-time inference speed of 148.3 FPS, a 22.5% reduction in parameter count, a 0.3% improvement in mAP, and a 15.0% decrease in GFLOPs. It achieves the lightweight design of the improved YOLO11 and the improvement of detection accuracy for forest fire targets.
Gao et al. (2026) studied this question.