Early Forest Fire Detection With UAV Image Fusion: A Novel Deep Learning Method Using Visible and Infrared Sensors
Kunlong Niu, Chongyang Wang, Jianhui Xu, Jianrong Liang, Xia Zhou, Kaixiang Wen, Minjian Lu, Chuanxun Yang
Abstract
Global warming has significantly increased the frequency of forest fires. Unmanned aerial vehicles (UAVs) provide rapid response and real-time monitoring, offering unique advantages over traditional human inspections and satellite monitoring. Their ability to monitor large forest areas during the early stages of fires supports timely warning. UAVs typically detect fires by capturing visible and infrared images. Visible images are effective for smoke detection but are influenced by environmental factors, while infrared images are better at detecting heat but can misidentify fires when the temperature difference between the fire and its surroundings is minimal. Additionally, challenges in image registration often occur when aligning the two image types for fusion. Therefore, this research proposes a novel method to early forest fire detection by fusing visible and infrared images and creating a dataset. The main contributions include: 1) the creation of a dataset containing 2752 synchronized visible and infrared image pairs to overcome existing dataset limitations; 2) the application of deep learning techniques to enhance image registration and fusion, incorporating an improved algorithm that increases automation; and 3) the development of the Forest Fire Detection Model—Fusion (FFDM-F) model, based on YOLOv5s and fused images, designed to accurately detect small fires at their early stages. The results show that the improved registration method effectively aligns visible and infrared images, optimizing the fusion process and enhancing the use of multisource information. Additionally, FFDM-F achieves over a 10% improvement in precision for small fire detection compared to traditional methods and reduces misidentifications associated with single-source images. This research contributes to multisource image fusion for forest fire detection, providing a more accurate and reliable early warning tool and laying the foundation for future work in this field.