Machine-learning-enabled three-dimensional firebrand transport characterisation: combustion enhancement, morphology and rotational dynamics

Firebrand spotting, responsible for 30–60% of structure ignitions in wildland–urban interface zones, has historically been constrained by slow, inter-operator-variable manual tracking. This work introduces an integrated stereo-vision and YOLOv8n-seg framework for automated three-dimensional firebrand transport quantification. The automated framework processes over 250,000 frames at 70 fps (≈2800 × speedup), achieves 95.9% mean average precision (mAP), and reduces per-trajectory analysis from 5 to 10 h of operator time to under 10 min. Position uncertainty is ±2.1 mm and velocity uncertainty ±0.15 m s⁻¹. The 108-experiment factorial dataset reveals that vortical flow modification increases path-integrated transport distance by ≈41% over the quiescent control, with active combustion adding a further ≈2% through effective buoyancy enhancement (Fbuoy = 1.15–1.24) and mass loss. Maximum transport heights rise by 9.96% under combined burning + vortex relative to vortex-only. Rotation-frequency and orientation-persistence statistics are consistent with rotation-mediated stabilisation for open-shaped firebrands, although passive aerodynamic alignment cannot be excluded with the present measurements. Morphology-specific drag coefficients (0.68–1.47) and effective buoyancy enhancement factors (1.15–1.24) provide laboratory-derived parameters for future wildfire spotting model development.