Farmland straw burning can produce incipient fire spots that occupy only a few pixels in UAV images and are easily confused with straw reflections, soil highlights, smoke, and illumination changes. This study proposes FireFly-YOLOv11, a real-time detector for tiny fire-spot detection under edge-device constraints. A stride-4 P2 detection head is added to the YOLOv11 prediction hierarchy to preserve fine spatial details for small targets. A Warm-Contrast Cue Attention (WCCA) module enhances fire-related saliency by jointly modeling local contrast variation and learnable warmth-inspired appearance cues. An Adaptive Asymmetric Atrous Feature Fusion (A3F) module adaptively aggregates multi-scale context while retaining low-level details through asymmetric gated residual fusion. Experiments on the StrawBurning UAV dataset show that FireFly-YOLOv11 achieves 88.6% Precision, 92.3% Recall, 91.7% mAP@0.5, and 41.0% mAP@0.5:0.95 at 50 FPS on an NVIDIA Jetson Orin Nano Super Developer Kit. Compared with baseline YOLOv11, it improves the four accuracy metrics by 2.1, 1.8, 2.7, and 2.0 percentage points. Ablation results confirm that P2, WCCA, and A3F provide complementary gains for UAV-based farmland fire monitoring.