Fire Detection In Wetland Using YOLOv4 And Deep Learning Architecture

One of the problems in wetlands is fire. Apart from material loss, land fires also cause vegetation degradation which has an impact on biodiversity and indirect effects such as haze. Current fire detection approach is utilizing Wireless Sensor Networks (WSN), but its use has disadvantages, namely the possibility of false alarms, limited space, and high-power consumption when using many sensors and additional cameras. Based on the shortcomings of WSN, several studies have suggested the use of Computer Vision. This is because the use of computer vision has advantages in terms of direct verification and only requires a camera device. One of the computer vision techniques is object detection, where fire and smoke are objects in an image. The current development of object detection that has good accuracy performance is YOLOv4. Therefore, YOLOv4 is used for detection and classification of wetland fires with image and video input. YOLOv4 model training using the transfer learning method resulted in 92.39 mAP performance on the CSPDarknet53-PAN-SPP architecture and experimental results on CSPResnet50-PAN-SPP (Full Mish Activation) resulted in mAP 92.13. In testing images and videos, CSPDarknet53-PAN-SPP has a higher accuracy result of 94% on images and 90% on video. The YOLOv4 average speed of the CSPResnet50-PAN-SPP model reaches a higher average value of 275.63 milliseconds in the image and 14.7 FPS in the video than CSPDarknet53-PAN-SPP. Therefore, the YOLOv4 model CSPDarknet53-PAN-SPP is better to serve as a prototype for detection and classification of fires in wetlands in term of accuracy.