UAV-based wildfire analysis

Abstract

Wildfire detection in the densely forested and remote regions of Finland poses significant challenges. Unmanned Aerial Systems (UAS) have transformed the monitoring and management of environmental hazards, including wildfires. Traditionally operated by human pilots, there is an increasing interest in deploying autonomous drone swarms to enhance efficiency and safety. This study introduces an innovative approach to wildfire detection using autonomous drones equipped with advanced sensory technologies such as RGB and thermal cameras. Additionally, the study introduces a dataset comprised of UAV-captured RGB and thermal video data, aiming to advance wildfire detection methodologies. It includes high-resolution images annotated through a semi-automatic method that leverages thermal information to annotate RGB images. The main goal of this study was to evaluate the utility of the dataset by applying established deep learning models, specifically ResNet and YOLO, and comparing their performance in unimodal and multimodal detection approaches. Intra-set evaluations on our novel dataset and inter-set evaluations through cross-validation with the Flame-1 and Flame-2 datasets demonstrate the robustness and applicability of our dataset in wildfire detection scenarios. The introduction of the UAV-RGBT dataset, along with the analysis conducted using novel deep-learning models, marks a significant advancement in wildfire management. This valuable resource has the potential to lead to cost-effective and environmentally sensitive solutions in remote sensing and emergency response strategies.