A UAV-Based Ignition Detection in Prescribed Fires Using Deep Learning ^†

Abstract

Forest wildfires have been an aggravating disaster during the past decades due to the rise of global temperatures. Forest management operations like prescribed fires are paramount for preserving these environments, although they present intrinsic risks. The FoCoR project aims to exploit UAVs with multispectral cameras to help manage prescribed fires by applying real-time detection and segmentation of ignitions. This paper proposes and details a basic supervised Deep Learning model capable of accurately detecting and segmenting prescribed fires. The model is based on the Mask R-CNN framework and is optimized with the best f1-score of approximately 70% which was considered a good starting point for further development. The used image dataset, consisting of more than 2500 polygonal labeled aerial RGB images acquired during prescribed fires will also be made publicly available for training more models in the future.