The Open Forest Science Journal
2008, 1 : 37-53Published online 2008 July 9. DOI: 10.2174/1874398600801010037
Publisher ID: TOFSCIJ-1-37
Joint Regional Simulation of Annual Area Burned in Canadian Forest Fires
ABSTRACT
Forest fires have a major impact on Canada’s forest carbon balance. National level simulations are used to gauge the potential impact of forest management, climate and other important factors on the carbon balance. In this process a joint simulation of the area burned in 29 Canadian forest fire regions is required. This study presents seven simulation algorithms for this complex task. Simulated areas burned are compared to historic data for 1959-1999 with respect to: 1) the average, 2) the variance, 3) the cumulative distribution function, 4) the first and last quartile, 5) skewness, and 6) kurtosis. Comparisons occur at three levels: regional, supra-regional, and combined regional level. Simulations based on a resampling of historic records (bootstrap) confirmed that the data were too sparse for this technique. Treating regions as independent (zero covariance) leads to a significant underestimation of the interannual variance of area burned at the supra regional and combined regional level. Thus an irregular and patchy interregional covariance structure must be taken into account. Only five regions could be viewed as independent from all other regions. A model-based simulation was challenged by the highly skewed and irregular regional distributions of area burned to which no known multivariate distribution would fit well. At the grouped regional and combined regional levels the best fit to historic data was achieved by first generating correlated probits with a multivariate-t distribution with four degrees of freedom, followed by a plug-in of these probits into empirical quantile interpolation functions. Approaches with time-varying or random covariance structures were also promising.
Correlationstratified bootstrapmultivariate normal