Prediction of Seasonal Climate in a Low-Dimensional Phase Space Derived from the Observed SST Forcing

Abstract

A methodology is presented to make an optimal use of the global SST for the prediction of seasonal climates. First, the space?time extended principal component analysis was applied to the key SST forcing regions, such as the tropical Pacific and the Atlantic, to establish a low-dimensional phase space model. This allows a nonlinear prediction, in terms of analogs found in the nearest neighborhood of the state associated with the initial time of prediction. Second, the predicted results derived independently from those different SST forcing regions are then linearly combined using the best linear unbiased estimates based on all available verification periods under a cross-validation scheme. This enables optimal use of the predictive skills inherent to each of the key SST forcing regions for each climate zone. The proposed methodology is justified by the analysis of the origins of predictive skills for seasonal predictions based on SST predictors (the geographical distribution of the skill scores and their time changes). Application was made to the prediction of winter (December?January?February) surface air temperatures over North America, based on the observed monthly mean data from January 1949 to December 1996. Significant skill scores were found over most parts of North America. The superiority of nonlinear prediction was demonstrated. It is concluded that the low-dimensional phase space approach may be used as an effective tool for seasonal forecasting.