An assessment of existing wildfire danger indices in comparison to one-class machine learning models

Abstract Predicting wildfires using Machine Learning models is relevant and essential to minimize wildfire threats to protect human lives and reduce significant property damage. Reliance on traditional wildfire indices for forecasting wildfires has failed to provide the expected prediction outcomes, resulting in limited application of these models. Thus, this research compares the outcome of wildfire forecasting using fire danger rating indices against Machine Learning model outcomes. Furthermore, the performance effectiveness of the fire danger rating indices and Machine Learning model outcomes are assessed using the same wildfire incidents. The One-class Machine Learning algorithms used are Support Vector Machine, Isolation Forest, Neural network-based Autoencoder and Variational Autoencoder models. The two global wildfire indices investigated were the US National Fire Danger Rating System for California and the McArthur Forest Fire Danger Index for Western Australia, using similar features. For the same data sets, the National Fire Danger Rating System and the McArthur Forest Fire Danger Index prediction outcomes were compared with Machine Learning model outcomes. Higher wildfire prediction accuracy was achieved by the One-class models, exceeding the performance of the two wildfire danger indices by at least 20%. The implications of our research findings have the potential to influence both these wildfire indices and state-of-the-art methods in wildfire prediction by proposing alternative ML methods to model the onset of wildfires.