Machine learning prediction of fine woody fuel consumption in surface fires burning in eucalypt forest fuels

Background: Accurate prediction of the woody debris consumed in wildfires is important for both wildland management and carbon accounting. Aims: We investigate the combustion factor (defined as the mean diameter reduction rate of the assumed cylindrical woody debris after fire) for fine woody debris (FWD) with pre-burn diameters ranging from 6 to 50 mm in free-spreading surface fires. Methods: Experiments were conducted in the CSIRO Pyrotron combustion wind tunnel facility (Canberra, Australia). A database of FWD consumption was constructed from experimental observations featuring 17 predictor variables. Machine learning models were applied to predict the FWD combustion factor. Key results: Pearson correlation coefficient analysis indicated that the FWD combustion factor exhibited highly significant negative correlations with smouldering duration, pre-burn diameter and tunnel axial position of FWD. Conclusions: We conclude that our combustion wind tunnel experimental approach captures the underpinning fire behaviour physics of FWD consumption well. A binary classification model using a support vector classifier demonstrated the best results for predicting FWD consumption, with an overall classification accuracy of 74%. A ridge regression model achieved a mean absolute error of approximately 9% for modelling FWD consumption. Implications: Our results illuminate possible options for controlling woody fuel consumption during managed fires in landscapes.