Specification of Wintertime North American Surface Temperature

Abstract

Abstract The extent to which wintertime North American surface temperature can be specified based on simultaneous sea surface temperature (SST) is quantified for the period 1982–98. The term specification indicates that the predictor and predictands are not lagged in time, as would be the case for true prediction. Four state-of-the-art general circulation models (GCMs) and linear empirical models with predictors derived from observations and dynamical models are considered. Predictors are derived from model hindcasts using principal component analysis (PCA), canonical correlation analysis (CCA), and discriminant analysis. The last technique has appeared in the climate literature, but its use in the present context appears new. A distinguishing feature of this paper is that several methods and models are compared in a common framework. The specification skill of GCMs for the period 1982–98 is statistically significant in the northwestern region near Washington State, British Columbia, and central Canada, with some local correlations exceeding 0.6. The specification skill of GCMs is comparable to, or better than, the skill of the best empirical model for the particular 17-yr period examined. No single specification strategy was found to improve the model hindcast skill in all cases. Predictors derived from discriminant analysis generally lead to larger skill than predictors based on PCA or CCA. The signal-to-noise ratio varies greatly among models and appears to be, if anything, inversely related to the specification skill when discriminants are used as predictors. Predictors based on 500-hPa geopotential height can lead to specification skill at least as good as predictors based on land surface temperature. Evidence is presented for the existence of at least two distinct dynamically predictable components of land surface temperature arising from two distinct “flavors” of SST anomalies associated with El Niño and La Niña.