LM‐YOLO: A Lightweight Multi‐Scale Enhanced Model for Forest Smoke Detection Using Unmanned Aerial Vehicles

ABSTRACT

Existing forest smoke detection models face limitations in recognizing small targets, achieving real‐time performance, and maintaining high inference efficiency on edge devices. To overcome these challenges, this study proposes a novel lightweight multi‐scale You Only Look Once (LM‐YOLO) model that enhances detection accuracy and real‐time capability while ensuring computational efficiency. The LM‐YOLO integrates the backbone feature extraction and multi‐scale fusion mechanisms of YOLOv12n. Specifically, we introduce the RFMBlock to enhance the fused representation of shallow and deep features, considering the visual characteristics of forest smoke. Meanwhile, the two‐path downsampling module achieves efficient downsampling with minimal spatial information loss. To further improve localization accuracy for small and blurred smoke regions, we employ the Shape‐IoU loss function. Extensive experiments on the public try123‐v4 forest fire smoke dataset demonstrate that, compared with the baseline YOLOv12n under the same experimental conditions, LM‐YOLO reduces the number of parameters by 34.5%, decreases the computational cost by 30.2%, and improves detection precision by 4.7%. In addition, the model achieves 92.7% precision on the public try7 dataset. These results outperform existing methods and highlight the strong adaptability of LM‐YOLO for edge deployment, ultimately providing reliable technical support for early forest fire warning using unmanned aerial vehicles.