Ensemble Canonical Correlation Prediction of Surface Temperature over the United States

Abstract

The ensemble canonical correlation (ECC) prediction method is used to predict summer (July?September) and winter (January?March) seasonal mean surface temperature (Tsurf) over the United States. The predictors are the global sea surface temperature (SST), sea level pressure over the Northern Hemisphere Tsurf, and soil moisture over the United States from one to two seasons lead, as well as the model outputs from the NCEP seasonal forecast model. The canonical correlation analysis (CCA) prediction is performed for each variable separately. The predicted Tsurf fields form an ensemble. The ensemble forecast is the weighted average of its members. Both the simple ensemble forecast and the superensemble forecast are tested. The simple ensemble mean is the equally weighted average of its members. The weighting function for the superensemble forecast is determined by linear regression analysis. Overall, both ensemble forecasts improve skill. On average, the superensemble gives the best performance. For summer, both ensemble forecasts improve skill substantially in comparison with the CCA forecasts based on the SST alone. Different variables recognize different forcing. They have forecast skills over different regions of the United States. Therefore, the ensemble forecasts are skillful. For summer, the leading SST modes that contribute to the sources of skill are associated with the long-term decadal trends, ENSO, and variability in the North Atlantic. In addition to SSTs, soil moisture in March?May also plays an important role in forecasting Tsurf in summer. For winter, SSTs in the tropical Pacific associated with the decadal and ENSO variability dominate the contribution.