Assessing the Effectiveness of YOLO Architectures for Smoke and Wildfire Detection

This paper presents a comprehensive evaluation of YOLO architectures for smoke and wildfire detection, including YOLOv5, YOLOv6, YOLOv7, YOLOv8, and YOLO-NAS. We aim to assess their effectiveness in early detection of wildfires. The Foggia dataset is used for this, and performance metrics such as Recall, Precision, F1-score, and mean Average Precision are employed. Our methodology trains each architecture for 300 epochs, focusing on recall for its relevance in this area. The 'best models' are evaluated on the Foggia test set and further tested with a challenging, custom-assembled dataset from independent online sources to assess real-world performance. Results show that YOLOv5, YOLOv7, and YOLOv8 exhibit a balanced performance across all metrics in both validation and testing. YOLOv6 performs slightly lower in recall during validation but achieves a good balance on testing. YOLO-NAS variants excel in recall, making them suitable for minimizing missed detections. However, precision performance is lower for YOLO-NAS models. Visual results demonstrate that top-performing models accurately identify most instances in the test set. However, they struggle with distant scenes and poor lighting conditions, occasionally detecting false positives. In favorable conditions, the models perform well in identifying relevant instances. We conclude that no single model excels in all aspects of smoke and wildfire detection. The choice of model depends on specific application requirements, considering accuracy, recall, and inference time. This research enriches the field of computer vision for smoke and wildfire detection, laying a foundation for system enhancements and serving as a basis for future research to optimize detection effectiveness.