A Markov Chain Monte Carlo approach to joint simulation of regional areas burned annually in Canadian forest fires

A simulation of regional and national forest carbon balance in Canada requires, due to regional correlations, a joint simulation of areas burned (BA) in regional fires. Regional correlations of BA are largely determined by concurrent years of relatively large (LF) and small fires (SF). A binary Markov Chain Monte Carlo procedure (MCMC) is constructed for forecasting regional LF(SF) status on an annual basis. For each forecast year the regional BA-value is obtained by a random draw from region-specific empirical quantile functions for LF and SF years. In the MCMC the conditional likelihood of a regional allocation of nLF* LF-years is maximized; whereby nLF* is drawn from a distribution fitted to LF(SF) classified data of area burned in 29 Canadian forest fire regions between 1959 and 1999. Regional allocation is governed by region-specific autologistic functions. MCMC results confirmed regional and national means and variances while regional correlations were generally somewhat lower than in the data.