Cross-Dataset Evaluation of YOLOv8 for Unmanned Aerial Vehicle Fire and Smoke Detection: Benchmark Contamination, Zero-Shot Transfer, and Onboard Deployment on a Low-Cost Airframe

Wildfires need a fast response, and a small unmanned aerial vehicle (UAV) carrying its own detector is an appealing way to find them early. Almost every published UAV fire detector is validated on a single corpus, which leaves open the question that matters to an operator: how much of the reported accuracy survives a change in scene. We trained unmodified YOLOv8n, YOLOv8s, and YOLOv8m on two public benchmarks that ship official test splits, D-Fire and the UAV subset of the Flame and Smoke Detection Dataset, under one fixed recipe with three seeds on D-Fire, then evaluated every checkpoint on both test splits with frozen weights. In domain, YOLOv8m reached 79.04 ± 0.21% mAP@0.5 on D-Fire, within 0.04 points of the published value for the same architecture on the same split, and 92.71% on FASDD_UAV. Moved across corpora, the same weights fell 38 and 65 points below a locally trained model, with fire degrading about twice as far as smoke. A perceptual-hash audit then found near-duplicates of training images in 46.0% and 90.7% of the two test sets; re-evaluating on the uncontaminated remainder costs 2.8 and 12.4 points, closes two-thirds of the apparent difficulty gap between the corpora, and leaves corrected transfer shortfalls near 26 and 62 points. Onboard, an F450 carrying a Pixhawk 2.4.8 and a Jetson Nano A02 flew and detected a controlled fire, but full-resolution inference ran 4.0 to 41.4 times slower than the video it consumed. Both gaps must close before a platform of this class can support operational wildfire monitoring.