Time Series Fuzzy Bayesian Logistic Regression for Wildfire Event Prediction

Abstract Wildfires are among the most destructive natural hazards, posing significant threats to ecosystems, communities, infrastructure, and economies. Accurately predicting the likelihood of wildfires enhances the effectiveness of preventive actions. Wildfire data is severely imbalanced toward the non‐fire observations due to the rare nature of wildfires. The features included in wildfire data cause separation problems and encounter seasonality. The methods used to predict wildfire probabilities need to account for these data characteristics. This study proposes a new method, TS‐FBLR, for predicting wildfire event probability. This method uses time series elements and fuzzy Bayesian modeling to handle class imbalance, separation, and seasonality in wildfire data. Extensive experimentation on 16 forests across Canada, Europe, and Australia, with data from the last 12 years, shows that TS‐FBLR outperforms 11 benchmarking methods for wildfire event classification. An out‐of‐sample benchmarking on data with a 12‐month forecast horizon demonstrates that TS‐FBLR accurately handles imbalance and separation problems with unseen data, improving implementation efficiency and providing real‐time predictions of wildfire probabilities.