IMPROVING SMALL FIRE TARGET DETECTION IN UAV IMAGERY: AN ENHANCED RT-DETR WITH MULTI-SCALE FUSION AND EXPERT ROUTING

Early fire detection is of paramount importance for forest fire prevention, yet traditional monitoring methods (e.g., satellites and ground-based stations) suffer from poor real-time performance or limited coverage. Unmanned aerial vehicles equipped with computer vision offer a novel solution for fire detection, but complex backgrounds, small flame and smoke targets, and varying illumination and weather conditions make accurate recognition challenging. In this work, we enhance the real-time detection Transformer model RT-DETR by designing a hybrid encoder architecture tailored for UAV fire imagery. Key improvements include the integration of an Adaptive Spatial Feature Fusion (ASFF) module to reconcile multi-scale feature inconsistencies; incorporation of Efficient Channel Attention (ECA) to strengthen channel-wise representations; replacement of the Transformer's fully connected feed-forward network with a Gated Mixture-of-Experts (MoE) structure to boost model capacity; and a multi-layer Transformer feature aggregation strategy. We evaluate the improved model on a UAV smoke fire dataset. Results show a significant uplift in both detection accuracy and recall: at an IoU threshold of 0.5, the enhanced RT-DETR achieves over 88.8% mAP—an approximate 2% gain over the original RT-DETR and superior performance compared to YOLO-series baselines. Ablation studies confirm that ASFF fusion, multi-attention mechanisms, and the MoE architecture each contribute meaningfully to small-target fire detection. Crucially, these advances incur negligible additional inference latency, enabling real-time intelligent monitoring for wildland fire scenarios.