YOLO-Based Sequential Image Approach for Improved Smoke Detection in Early-Stage Wildfires

Abstract

Wildfires in northern Thailand, particularly during the dry season from November to May, are a major source of hazardous air pollution and sharp increases in PM2.5 levels. Despite no-burning policies, conventional monitoring methods such as lookout towers remain inadequate due to limited visibility, staffing shortages, and delayed response times. Existing smoke detection models mostly rely on single-frame analysis, which often misclassifies fog or clouds as smoke and fails to detect faint or obscured smoke plumes. This reduces the reliability of earlywarning systems in real-world deployment. The objective of this study is to evaluate whether time-series analysis can improve detection performance over single-frame baselines. We fine-tuned a YOLOv8n model on the FireSpot dataset and designed a sequential evaluation protocol that aggregates predictions across short image sequences, converting per-frame outputs into eventlevel classifications. This temporal consistency filters out transient false positives while preserving true smoke evidence. Experimental results show that per-image classification achieved a precision of 0.9780, a recall of 0.8230, an F1-score of 0.8940, and a balanced accuracy of 0.9020. The sequential framework improved performance with perfect precision of 1.0000, a recall of 0.8970, an F1-score of 0.9460, and balanced accuracy of 0.9490. These results confirm that temporal modeling significantly enhances robustness and reliability, providing a practical pathway for early-warning deployment in wildfire-prone regions.