Pyramid Attention Based Early Forest Fire Detection Using UAV Imagery

Abstract

Early forest fire detection using unmanned aerial vehicle (UAV) imagery is widely implemented in recent years. However, as the size of moving object in UAV imagery changes dramatically, the issues of focusing more informative areas of an image, large-scale variation as well as temporal consistency preservation have yet to be resolved due to the wide field of vision of UAVs in flight. To address these issues, in this work, we concerned the visual characteristics of early stage forest fire in UAV imagery and employ pyramid attention mechanism. We engineered a novel convolutional neural network (CNN) based on the state-of-the-art backbone by stacking our proposed PyrAtten blocks. We built a large dataset of early stage forest fire, including real world aerial photographs shot using our UAV prototype. Extensive evaluation of PyrAtten demonstrates its efficiency and effectiveness in detecting multi-scale objects of fire and smoke in UAV imagery, even if the object takes up a very small proportion of the whole image. Compared with existing similarly-sized networks, average detecting accuracy of PyrAtten-ResNet-50 reaches 97.5%, with negligible increase in computational overhead.