Advanced wildfire detection using generative adversarial network-based augmented datasets and weakly supervised object localization

Abstract

Owing to abnormal climate phenomena worldwide, forests are becoming dry and heat waves have started to occur, increasing the damage caused by wildfires. In addition to causing significant human and material damage, wildfires are also a major cause of critical pollutant emissions, in which fine dust generated by incomplete combustion pollutes the atmosphere, soil, and water. Early detection and monitoring are some of the main ways for minimizing wildfire damage, and a topic of research interest in various fields of artificial intelligence and computer vision. However, the lack of wildfire occurred image datasets is still challenge. Training deep learning model in this environment, can lead mis-detection when burning point is far from the camera or according to objects similar to flame and smoke. Our study attempted to create synthetic wildfire images in various shapes by inserting damage into a free-wildfire image using generative adversarial network (GAN) and Weakly supervised object localization (WSOL). The synthesized image can used as training data for object detection by applying the WSOL method with gradient-weighted activation map (Grad-CAM). Additionally, the YOLOv5s model was improved by adding a channel attention module; sequence-and-excitation (SE) layer and replace loss function as CIoU to address the issue of wildfire false detection in fire-like object and miss detection in small size smoke. Our proposed method, produced results as high as 7.19% in F1-score and 6.41% in average precision (AP) when compared to the existing traditional method. To use a deep learning model in practice, a lightweight model should be applied to the embedded models while maintaining high performance. The developed AI model was applied to the established drone and CCTV-based wildfire monitoring system, and a virtual experiment was conducted by generating virtual wildfires near forests in Korea. • It has steadily expanded with the advancement of deep learning in computer vision, but there is a problem that makes it challenging to build a robust model for a variety of environments due to data scarcity, misdetected objects that look like smoke or flames, and small objects of which burning point is too far from the camera. • SE block and CIoU was employed to address misdetection to expect that SE can extract important feature in smoke/fire of irregular shapes and colors, that CIoU can promoting the positioning accuracy of the small object. • According to test results proposed method can have sufficient accuracy comparing to SOTA model in real time object detection. • Our proposed method can be expected to contribute to minimizing damage to global environmental disasters caused by wildfires by ensuring accuracy and speed in Early wildfire detection than exist model.