YOLOv8-CNR: a lightweight fire and smoke detection model for forest fire scenarios based on improved YOLOv8

Introduction To address irregular morphology and drastic scale variations of flame and smoke targets in forest fire detection, and the difficulty of lightweight models in balancing accuracy and real-time performance, this paper proposes YOLOv8-CNR based on improved YOLOv8n. Methods The model optimizes backbone, neck, and loss function: C3_ConvNeXtv2 enhances flame texture and smoke edge representation; RepNCSPELAN4 strengthens multi-scale flow and small-target smoke features; SIoU improves localization of irregular targets. Results On M4SFWD with its trajectory-level split, YOLOv8-CNR achieves 79.8% mAP50 and 39.4% mAP50-95 with 2.52 M parameters, improving 7.5 and 2.6 percentage points over tuned YOLOv8n, at 238.1 FPS. This split eliminates near-duplicate frame leakage across subsets. Because the test set consists entirely of six Scene 4 camera runs that are absent from both training and validation, the M4SFWD results are interpreted as controlled comparisons under a scene-held-out setting rather than same-distribution generalization. Cross-dataset transfer is further evaluated through zero-shot transfer: trained only on M4SFWD, the model achieves 78.3 and 82.7% mAP50 on SMOKE_dataset and FLAME_dataset, respectively. Retrained on these datasets, it achieves 90.2 and 90.4% mAP50, demonstrating favorable real-world applicability. Ablation experiments show complementary and partially overlapping effects among the three modules, with the complete model achieving a 2.6 percentage point improvement in mAP50-95 over the baseline. Discussion The proposed model improves overall and smoke-related precision and recall, while flame precision slightly decreases (81.1% vs. 82.4%), flame mAP50-95 increases from 42.0 to 45.2%. This method provides an efficient solution for real-time flame and smoke detection in complex forest scenarios.