Spring predictability barrier of ENSO events from the perspective of an ensemble prediction system

Based on an ENSO (El Nino-Southern Oscillation) ensemble prediction system (EPS), the seasonal variations in the predictability of ENSO are examined in both a deterministic and a probabilistic sense. For the deterministic prediction skills, the skills of the ensemble-mean are sensitive to the month in which the forecast was initiated. The anomaly correlations decrease rapidly during the Northern Hemisphere (NH) spring, and the root mean square (RMS) errors have the largest values and the fastest growth rates initialized before and during the NH spring. However, the probabilistic predictions based on the verification methods of the relative operating character (ROC) curve and area both show that there are no strong seasonal variations for the two extreme (warm and cold) ENSO events. For the near-normal events, the seasonal variations of the probabilistic skills are much more obvious, and the ROC areas of the ensemble forecasts made in the spring are clearly smaller than those of the ensemble forecasts that began during other seasons. At the same time, the probabilistic prediction skills of the EPS for all three events that only consider the initial perturbations are also clearly sensitive to the initial months. This was indicated by the fact that the most rapid decrease of the ROC area skill occurs as the hindcasts proceed through the spring season. A further signal-to-noise ratio analysis reveals that potential sources of the predictability barrier in the probabilistic skills for the EPS are namely that the spring is the period when stochastic initial error effects can be expected to strongly degrade forecast skill, and that small predicted signals can render the system noisier by further limiting the predictability. However, reasonable considerations of the model-error perturbations during the ensemble forecast process can alleviate the barrier caused by initial uncertainties through coordinately simulating the seasonal variations of the forecast uncertainty in order to significantly improve the probabilistic prediction skills and then to disorder the seasonal predictability related to the SPB. (C) 2010 Elsevier B.V. All rights reserved.