Enhancing UAV-Based Fire Detection via Adaptive Anchor Clustering and Lightweight Bidirectional Feature Fusion

Early detection of fires is important for reducing casualties and property damage; however, existing vision-based methods show reduced performance in complex scenes involving small targets and heterogeneous data distributions. This paper proposes an adaptive fire detection framework based on YOLOv11, which improves anchor box configuration and multi-scale feature fusion. Specifically, a K-means–based adaptive anchor clustering strategy is used to generate anchor boxes according to the scale distribution of flame and smoke targets. In addition, a lightweight bidirectional feature pyramid network (BiFPN) with depthwise separable convolutions is introduced to improve feature representation for small and sparse targets while maintaining low computational cost. Experiments on the dataset show that the proposed method achieves 78.443% mAP 50 , 56.327% mAP 50–95 , and 70.019% Recall, corresponding to improvements of 4.131%, 3.646%, and 5.402% over the baseline YOLOv11. Ablation and cross-dataset experiments further show that each component contributes to performance improvement, particularly in small-target and multi-scale scenarios. These results indicate that the proposed method improves detection performance under complex fire scenarios.