ENSO ensemble prediction and predictability for the past 148 years from 1856--2003

Several important issues of El Nino-Southern Oscillation (ENSO) predictability were studied using the latest version of the Zebiak-Cane model, singular vector (SV) analysis, ensemble hindcast, and information theory for the period of 148 years, e.g., the dominant factors controlling ENSO prediction skills, the useful precursors of forecast skill, ensemble construction and probabilistic verification.More precisely, there are four main sections in this thesis. (1) A fully physically-based tangent linear model was constructed for the Zebiak-Cane model and a singular vector (SV) analysis for the 148 year (1856-2003) was performed. It was found that the leading SVs are less sensitive to initial conditions while singular values and final perturbation patterns exhibit a strong sensitivity to initial conditions. The dynamical diagnosis shows that the total linear and nonlinear heating terms play opposite roles in controlling the optimal perturbation growth. (2) Relationships between the singular values and actual prediction skill measures were investigated. At decadal/interdecadal time scales, an inverse relationship exists between the leading singular value (S1) and correlation-based skill measures whereas an in-phase relationship exists between the S1 and MSE-based skill measures. However, S1 is not a good predictor of prediction skill at shorter time scales and for individual predictions. An offsetting effect was found between linear and nonlinear perturbation growth rates, which have opposite contributions to the S1. (3) Ensemble and probabilistic ENSO predictions were performed for the 148 yrs. Four typical ensemble construction strategies were investigated. Results suggest that "reliability" is more sensitive to choice of ensemble construction strategy than "resolution". The fourth strategy produces the most reliable and skillful ENSO probabilistic prediction, benefiting from the contribution of the stochastic optimal winds and singular vector of SSTA. (4) Information and ensemble-based potential predictability measures are explored on multiple time scales. Relative entropy is better than predictive information (PI) and predictive power (PP) in quantifying the correlation-based prediction skill whereas PI/PP is a better indicator in estimating mean square error (MSE)-based prediction skill.