Statistical Forecasts of Western Wildfire Season Severity

Strong associations between observed climate anomalies at lags of one season to years in advance and normalized area burned in the western U.S. wildfire season have recently been described using a newly compiled comprehensive gridded western regional fire history (Westerling et al 2001b). Earlier studies of fire scar dendrochronologies and local fire histories have demonstrated that large-scale climate patterns are linked to the severity of the wildfire season in various regions of the U.S. at similar lead times (Simard et al.1985, Swetnam and Betancourt 1990, Balling et al 1992, Swetnam and Betancourt 1998, Jones et al. 1999). These relationships and the availability of a comprehensive western wildfire history motivate this experimental statistical forecast methodology for the western wildfire season. Previous work (Westerling et al 2001a & b) has established that lags of the Palmer Drought Severity Index can be used to forecast normalized acres burned at lead times of a season to years in advance, using co-located PDSI values as regressors. This work has also shown that regional indices describing modes of variability in PDSI—represented by leading principal components (PCs) of lagged PDSI values—show similar skill in forecasting western wildfire season severity. Moreover, models based on these regional indices show impressive predictive skill even in locations where strong associations between local PDSI values and normalized acres burned are lacking. Canonical Correlation Analysis (CCA) offers a method for constructing western wildfire season severity models whose prediction skill derives from spatial and temporal patterns in climate spanning the western U.S. In this example