Wildfire and Smoke Detection Using YOLO-NAS

Wildfires pose a persistent threat to the environment, human lives, and property globally, necessitating effective early detection for firefighting, environmental protection, and public safety. This paper is focusing on the YOLO-NAS model and the Adam/AdamW optimizers for accurate and swift wildfire and smoke detection in various scenarios. The motivation lies in the critical need for a versatile detection model to address the challenges posed by these devastating natural disasters and their environmental impacts. The main contribution of this study resides in the integration of data from aerial, infrared, satellite, and web cam sources into a unified dataset with YOLO-NAS and the evaluation of the Adam W optimizer which was relatively unexplored in previous papers. The YOLO-NAS-M architecture, coupled with the Adam optimizer, is chosen as it provided the best global performance on evaluation when compared to other models. The model demonstrates improved performance in mean average precision and recall across diverse datasets. It achieves a mAP@0.50 of 73.39% and a mAP@0.50:0.95 of 42.86% on the full dataset. For the webcam and infrared datasets, mAP@0.50 is over 91 %, highlighting its potential for real world applications. Despite challenges in precision and F1 score, the study contributes valuable insights to wildfire detection, emphasizing the importance of architectural choices and optimization strategies.