Statistical Model for Forecasting Monthly Large Wildfire Events in Western United States

Abstract

The ability to forecast the number and location of large wildfire events (with specified confidence bounds) is important to fire managers attempting to allocate and distribute suppression efforts during severe fire seasons. This paper describes the development of a statistical model for assessing the forecasting skills of fire-danger predictors and producing 1-month-ahead wildfire-danger probabilities in the western United States. The method is based on logistic regression techniques with spline functions to accommodate nonlinear relationships between fire-danger predictors and probability of large fire events. Estimates were based on 25 yr of historic fire occurrence data (1980?2004). The model using the predictors monthly average temperature, and lagged Palmer drought severity index demonstrated significant improvement in forecasting skill over historic frequencies (persistence forecasts) of large fire events. The statistical models were particularly amenable to model evaluation and production of probability-based fire-danger maps with prespecified precisions. For example, during the 25 yr of the study for the month of July, an area greater than 400 ha burned in 3% of locations where the model forecast was low; 11% of locations where the forecast was moderate; and 76% of locations where the forecast was extreme. The statistical techniques may be used to assess the skill of forecast fire-danger indices developed at other temporal or spatial scales.