UAV-Based Efficient and Reliable Wildfire Detection on the Edge

Wildfires have become an increasingly frequent and destructive force, necessitating the development of innovative monitoring and detection systems. In this paper, we explore the use of unmanned aerial vehicles (UAVs), commonly known as drones, equipped with advanced computer vision technology for real-time wildfire detection. Leveraging the flame2 dataset, which includes multispectral imagery combining RGB and infrared (IR) images, this paper employs a modified YOLOv8 (You Only Look Once) convolutional neural network to detect fires in various environmental conditions. Empirical evaluations using flame 2 demonstrate a significant performance improvement over state-of-the-art wildfire detection systems. Our model attains an impressive $\mathbf{9 9. 5 8 \%}$ accuracy and F 1 -score of $\mathbf{9 9. 7 6 \%}$ on wildfire classification at heterogeneous conditions (smoke, night-time, among others). It can accurately detect precised fire regions with higher confidence demonstrating the reliability of the fire detection system.