Short-term forecasting of human-caused wildland fire occurrence: a case study for two regions in Ontario, Canada

Wildfire occurrence prediction (FOP) models are used as decision support tools by fire management agencies. However, there appears to be limited research on making predictions beyond the current day, especially for short-term forecast horizons. Here, we consider short-term forecasting of FOP one to four days into the future using data from the Fort Frances and Dryden Fire Management Districts in the Province of Ontario, Canada. Forecasts of Fine Fuel Moisture Code, when used as the predictors for fire occurrence, are subject to prediction errors. Ignoring such errors would lead to biased or incorrect estimators. We show that the forecasting error distributions can be modeled by normal mixtures. Then we employ a modified simulation-extrapolation (SIMEX) algorithm that can incorporate these normal mixture models when developing short-term forecasting models for FOP. Since we are interested in predicting both the presence and the number of fires, and to demonstrate that this method can be employed to forecast both the presence of at least one fire and for counts of fire activity, both logistic- and count-based models are developed. Through simulation, we show that ignoring measurement errors leads to biased estimators, whereas the SIMEX estimators mitigate this issue.