Partial least squares regression as a useful tool for determining the predictable component of atmospheric fields

A common goal in atmospheric and oceanic sciences is to diagnose the relationships among different atmospheric and oceanic fields for the purposes of prediction, hypothesis testing, and exploratory analysis. However, most conventional approaches rely on ad hoc indices to represent spatial field variability or decomposition methods with constraints that limit their use for prediction problems. Here we propose a method that combines empirical orthogonal function analysis with partial least squares (PLS) regression to estimate the component of the atmospheric circulation that is predictable by a predictor field of interest. We test this approach with an atmospheric general circulation model forced by the observed, time varying sea surface temperatures. This test demonstrates that the PLS regression approach is generally quite successful in reproducing the predictable component, i.e., ensemble mean response, of atmospheric fields with the record of only a single ensemble member. Thus, the approach suggests promise for the analysis of observational data and the potential use in extended-range forecasts, climate field reconstructions, and diagnostic analyses.