Methodology for detecting firebrand generation and the analysis of influential variables in quantification

Firebrands play a dominant role in wildland fire propagation, especially in spotting, where embers are carried away from the main fire to ignite new areas. Three mechanisms govern their dynamics: generation, transport, and landing. While transport and landing have received more attention in the literature, generation remains poorly understood due to challenges in real-time quantification and detection. This shortcoming undermines fire spread models that require accurate estimation of particle properties and numbers. To address this, a UNet-based convolutional neural network (CNN) image processing method was developed to detect and segment firebrands from images and video. The model demonstrated over 80 % accuracy with limited training data, suggesting potential for field use with low-cost, heterogeneous imaging systems. Experiments were conducted in large wind tunnel (1.5m × 2.1m × 6 m) at uniform wind speeds up to 6.2 m/s and test velocities between 0.5 to 1.5 m/s. Douglas Fir samples were burned, and mass loss was measured using balances. Fuel positioning affected firebrand production: free samples (1.5 kg) exhibited multiple intense burning stages due to deformation, whereas restrained samples (0.76 kg) burned more uniformly. The proposed CNN approach offers a promising tool for supporting firebrand detection.