YOLO-MC: An algorithm for early forest fire recognition based on drone image

Abstract Forests not only provide important resources for the earth, but also play an important role in maintaining ecological balance. Once forest fires occur, they will pose a serious threat to human beings. As a convenient and high-definition remote sensing equipment, unmanned aerial vehicle provides effective support for forest fire monitoring. The early forest fire smoke flame target is small and the dynamic change is difficult to identify. In this study, a forest fire dataset is first constructed, and an early forest fire detection model, YOLO-MC, based on YOLOV8, is proposed. Lightweight backbone network MobileNetV3 is used to replace the original backbone. Convolutional block attention module attention mechanism is added to make the model pay more attention to local information connection. Bidirectional feature pyramid network and small target layer are introduced to dynamically assign weights to identify small targets. Dynamic head is used to improve the flexibility of model inspection. Generalized intersection over union is used as a loss function to improve the accuracy of target positioning. Compared with the benchmark model, the accuracy and average accuracy of YOLO-MC are increased by 4.3 and 3.6%, respectively, and the number of parameters is reduced by 2 × 10 4 , which indicates that YOLO-MC can improve the identification efficiency of early forest fires and realize the early detection of forest fires at the same time. It helps in identifying forest fires as early as possible and has certain practical significance.