Tree-Structured Modeling of the Relationship Between Great Lakes Ice Cover and Atmospheric Circulation Patterns

Seasonal maximum ice concentration (percentage of lake surface covered by ice) for the entire Laurentian Great Lakes and for each Great Lake separately is modeled using atmospheric teleconnection indices. Two methods, Linear Regression (LR) and Classification and Regression Trees (CART), are used to develop empirical models of the interannual variations of maximum ice cover. Thirty-four winter seasons between 1963 and 1998 and nine teleconnection indices were used in the analysis. The ice cover characteristics were different for each Great Lake. The ice cover data lent itself better to CART analysis, because CART does not require a priori assumptions about data distributions characteristics to perform well. The stepwise LR models needed more variables, and in general, did not explain as much of the variance as the CART models. Two variables, the Multivariate ENSO index and Tropical/Northern Hemisphere index, explained much of the interannual variations in ice cover in the CART models. Composite atmospheric circulation patterns for threshold values of these two indices were found to be associated with above-and below-normal ice cover in the Great Lakes. Thus, CART also provided insight into physical mechanisms (atmospheric circulation characteristics) underlying the statistical relationships identified in the models.