Abstract
Background Quantifying and predicting wildland fire behavior is crucial for fire management, ecological research and mitigating wildfire impacts. Rate of spread (ROS), fireline intensity (FI) and fire radiative power (FRP) are key fire behavior metrics.Aims This study leverages uncrewed aircraft systems (UASs) equipped with thermal infrared (TIR) sensors and machine learning models to quantify and predict fire behavior.Methods Using repeat-pass UAS-based TIR imagery, we derived high-resolution FRP, FI and ROS estimates and trained artificial neural network (ANN) and random forest (RF) models to predict ROS.Key results This approach predicted ROS with low error (mean absolute errors (MAEs) below 0.04 m s-1, root mean squared errors (RMSEs) below 0.06 m s-1 and R2 values above 0.90) in short-term predictions for a single prescribed grassland fire, while maintaining computational efficiency.Conclusions Both ANN and RF models performed well, but RF performed better, with less training data, lower propensity for overfitting and less sensitivity to spatial autocorrelation.Implications Although currently demonstrated as a proof of concept at a single site with a specific fuel type and short-term prediction horizon, our integrated methodology shows research and development potential for supporting data-driven wildfire management strategies aimed at mitigating fire impacts, optimizing resource allocation and improving firefighter safety.