An attention-enhanced YOLO framework for robust forest and land fire detection

Early detection of forest and land fires is essential for reducing environmental damage, economic losses, and risks to human safety. Recent advances in deep learning have enabled automated fire and smoke detection using real-time object detection frameworks. However, detecting small flames and diffuse smoke remains challenging due to complex backgrounds, low contrast, and varying environmental conditions. This study investigates the effectiveness of integrating the Convolutional Block Attention Module (CBAM) into YOLOv13 for forest and land fire detection. Three CBAM insertion strategies, namely backbone, neck, and detection head integration, were evaluated using the Fire-Smoke dataset under identical training conditions. Model performance was assessed using precision, recall, mean average precision at an intersection-over-union threshold of 0.5 (mAP50), and mean average precision across thresholds from 0.5 to 0.95 (mAP50–95). Computational efficiency was evaluated using parameter count, floating-point operations (FLOPs), inference latency, and frames per second (FPS). The results show that backbone-level CBAM integration achieved the highest detection performance among the evaluated configurations, reaching 86.5% precision, 82.5% recall, 87.7% mAP50, and 59.4% mAP50–95, outperforming the baseline YOLOv13s model. However, repeated-run validation indicates that the magnitude of improvement did not reach statistical significance at the α = 0.05 level, and the conclusions drawn should be interpreted within the scope of the dataset and experimental conditions used in this study. These findings suggest that lightweight attention-based feature refinement may contribute to improved wildfire detection performance while preserving the computational efficiency required for practical real-time monitoring applications.