Research On UAV-Based Forest Fire Detection

Forests not only regulate the climate and maintain the ecological balance,but also can contain water and mitigate the greenhouse effect.With an area of about 220.45 million hectares and a coverage rate of 23%,China’s forests are one of the important natural resources.Due to the high content of combustible materials in forest areas,forest fires can cause huge ecological and economic losses once they occur.And the initial detection of forest fires is crucial for fire prevention.Compared with satellite and aircraft detection methods,UAVs have become an important supplementary means of forest fire detection by virtue of their flexibility and economic advantages.In this paper,the key technologies of forest fire detection by UAVs are studied around:(1)UAV path planning research based on energy consumption.According to the actual forest fire detection needs,the operation area is selected and the path planning is carried out for it.Energy consumption is a key factor to be considered in UAV operations.According to the different energy consumption of UAVs in different flight states,this paper studies the total length of UAV flight paths,the number of turns and the optimal speed at different straight-line flight distances,and comes up with a law to determine the low energy consumption flight direction and the optimal speed.Finally,a low-energy path is planned and experimentally verified for the selected operation area.(2)Smoke detection algorithm research.The smoke characteristics are obvious in the early stage of forest fires,so smoke is used for fire determination.Using self-built smoke dataset for mainstream target detection algorithms Faster R-CNN,SSD,YOLOv5 training,the average detection accuracy is 84.5%,83.9%,84.7%,respectively,so YOLOv5 is selected as the detection model.However,the model cannot meet the real-time and high-precision requirements of UAV-based smoke detection,and should be further optimized.It is proposed to use Ghost Net,a lightweight network,to replace the YOLOv5 backbone network and fuse the attention mechanism,and to use C3 Ghost to replace the C3 module in Head.The improved YOLOv5-GCA model has 51.35% fewer parameters,1.2% higher detection accuracy,and 21.2% higher speed.Finally,ablation experiments and comparison tests are set up to demonstrate the superiority of the algorithm improvement.(3)Research on fire point detection and localization based on infrared images.When a forest fire occurs,the temperature of the area will rise sharply,and the fire point feature is more obvious in the infrared image.Using the self-built fire point IR dataset trained on the mainstream target detection algorithms Faster R-CNN,SSD,and YOLOv5,the average detection accuracy is 77.6%,77.1%,and 78.4%,respectively,and the performance is generally low.In YOLOv5 Head,the C3 module is replaced with C3 Ghost,and Concat is replaced with Bi FPN＿Concat,and then the YOLOv5 backbone network is replaced with Ghost Net.The improved YOLOv5-GB model improves detection accuracy by 3.1% and recall by 1.2%.Ablation experiments are set up to demonstrate the effectiveness of the improvement,and comparison experiments are conducted with the YOLOv5-GCA detection model as well as other models.When a fire is detected,the location of the fire point is calculated by converting between coordinate systems with the help of the latitude and longitude of the location where the UAV is located.(4)Design and development of forest fire detection system.In order to verify the performance of the algorithm studied in this paper,a forest fire detection system is designed and developed.Considering that some forest areas will adopt the installation of cameras to monitor whether fires occur in forest areas,the system is designed into two major functional modules: the camera-based video forest fire detection module realizes the smoke detection and fire point detection functions;the second UAV-based forest fire detection module system realizes the smoke detection,fire point detection,UAV fire point location and alarm functions.