Bivariate extreme value modeling of wildland fire area and duration

Abstract

Area burned and fire duration as two aspects of large wildland fires were assessed by extreme value statistics in this study. A data set of 64,474 fire records from Mississippi between January 1991 and December 2007 was used in the assessment. Large fires occurred mainly between February and May, in the Southeastern Region, and started as debris and incendiary fires. Area burned showed more extremal properties than fire duration. A positive and moderate association existed between area and duration of wildland fires. In the univariate analyses, the generalized Pareto distribution with a heavy tail characterized both the area burned and fire duration well. The bivariate extreme value analyses found that the asymmetric negative logistic distribution provided the best fit to the data. Incorporating the dependence between area and duration into the bivariate extreme value models and forecast of return levels generated more conservative estimates than those from the univariate analyses. The techniques adopted in this study can be applied to analogous data sets for other regions. The information on fire patterns and extremal behavior could be beneficial to forest planning and management. © 2013 by the Society of American Foresters.