Low-Cost UAV-Based Wildfire Smoke Detection

Abstract

Early wildfire detection is important because fires become more difficult and dangerous to control as they grow. This paper examines whether a lightweight computer-vision model could support low-cost wildfire-smoke monitoring using small unmanned aerial vehicles. The project was motivated by a Crazyflie drone equipped with an AI Deck, although the model was trained and tested using the publicly available Boreal Forest Fire dataset rather than images collected directly by the drone. The dataset contained 4,954 labeled UAV images from four forest locations in Finland. Four YOLO11n models were trained using a leave-one-location-out approach. In each experiment, images from three locations were used for training and images from the fourth, previously unseen location were used for testing. At a resolution of $640 \times 360$, the models achieved overall smoke-detection scores between 76.0\% and 93.0\%. This shows that the models could detect smoke in new forest environments, although performance varied between locations. The paper also examined the effect of image resolution and analyzed false detections caused by features such as fog, clouds, haze, and bright backgrounds. The results provide a proof of concept for future wildfire smoke detection using low-cost, resource-constrained drones. The source code, trained models, annotations, evaluation results, and reproducibility materials are available in the project's \href{https://github.com/krishnanroshan/Wildfire-Smoke-Detection-with-Location-Based-YOLO-Evaluation}{GitHub repository}.