Improved reliability of ENSO hindcasts with multi-ocean analyses ensemble initialization

Abstract

Currently, ensemble seasonal forecasts using a single model with multiple perturbed initial conditions generally suffer from an “overconfidence” problem, i.e., the ensemble evolves such that the spread among members is small, compared to the magnitude of the mean error. This has motivated the use of a multi-model ensemble (MME), a technique that aims at sampling the structural uncertainty in the forecasting system. Here we investigate how the structural uncertainty in the ocean initial conditions impacts the reliability in seasonal forecasts, by using a new ensemble generation method to be referred to as the multiple-ocean analysis ensemble (MAE) initialization. In the MAE method, multiple ocean analyses are used to build an ensemble of ocean initial states, thus sampling structural uncertainties in oceanic initial conditions (OIC) originating from errors in the ocean model, the forcing flux, and the measurements, especially in areas and times of insufficient observations, as well as from the dependence on data assimilation methods. The merit of MAE initialization is demonstrated by the improved El Niño and the Southern Oscillation (ENSO) forecasting reliability. In particular, compared with the atmospheric perturbation or lagged ensemble approaches, the MAE initialization more effectively enhances ensemble dispersion in ENSO forecasting. A quantitative probabilistic measure of reliability also indicates that the MAE method performs better in forecasting all three (warm, neutral and cold) categories of ENSO events. In addition to improving seasonal forecasts, the MAE strategy may be used to identify the characteristics of the current structural uncertainty and as guidance for improving the observational network and assimilation strategy. Moreover, although the MAE method is not expected to totally correct the overconfidence of seasonal forecasts, our results demonstrate that OIC uncertainty is one of the major sources of forecast overconfidence, and suggest that the MAE is an essential component of an MME system.

Appendix

1.1 Four ocean analyses used in hindcast AP/MAE

In hindcast MAE, four different ocean analyses were used as OICs, with two from ECMWF and two from NCEP. They are the ECMWF COMBINE-NV (Balmaseda et al. 2010), the ECMWF Ocean Reanalysis System 3 (ORA-S3; Balmaseda et al. 2008), the NCEP Climate Forecast System Reanalysis (CFSR; Saha et al. 2010), and the NCEP Global Ocean Data Assimilation System (GODAS; Behringer 2005). The GODAS and CFSR (ORA-S3) ocean analyses have been used to initialize the operational seasonal predictions made by NCEP (ECMWF). COMBINE-NV, which is only slightly different from ECMWF Ocean Reanalysis System 4 (ORA-S4), has been used to initialize their decadal predictions at ECMWF. Table 1 briefly summarizes the major characteristics of these ocean analyses, including the models, resolutions, assimilation methods, and the assimilated data. To show the systematic differences among the four ocean analyses, Figs. 5 and 6 present the standard deviation of SST/HC300 systematic differences among them and the signal versus noise ratio maps, respectively. From Fig. 5, it can be seen that SST shows little difference among them in the tropics expect for the far eastern coastal regions. There is larger difference among them in the subsurface as shown in Fig. 5b. Particularly, in the off-equatorial regions the difference in HC300 is larger than 0.2 °C, with less difference along the equator. Figure 6 also indicates that SSTs have relatively lower level of noises, while HC300s have larger level of noises, particularly over the off-equatorial regions.

Fig. 5

The standard deviation of the systematic differences in a SST and b the upper 300 m mean ocean temperature (HC300) for 1982–2007 among the four ocean analyses used in the study (i.e., COMBINE-NV, ORA-S3, CFSR, GODAS). Contours 0.2, 0.4, 0.6, 1, 15 are shown. The systematic differences are calculated as the difference between four individual datasets and their ensemble mean. Unit: °C

Fig. 6

The global maps of the signal versus noise ratio of a SST and b HC300 for 1982–2007 derived from the four ocean analyses (i.e., COMBINE-NV, ORA-S3, CFSR, GODAS). Contours 1, 2, 5, 10, 15 are shown. Here signal is defined as the interannual variance in the ensemble mean of four datasets; noise is defined as the variance of the difference between four individual datasets and their ensemble mean

Table 1 Brief summary of the used ocean analyses

1.2 Experimental design for hindcast AP/MAE

For hindcast AP_cbn, AP_ora3, AP_cfsr, and AP_gds, the atmosphere, land and sea ice initial states are specified in the same way, using the instantaneous fields from the CFSR. For each hindcast, four ensemble members are generated that differ in their atmosphere/land surface conditions, which are the instantaneous fields from 00Z of the first four days in April in the CFSR, respectively. For the OIC, to reduce the potentially negative effects of the mean biases in ocean analyses and the forecast model, and to make the predictions using OIC from different analyses comparable, we applied an anomaly ocean initialization strategy (e.g., Schneider et al. 1999) in these experiments. For this purpose, a monthly climatology for the CFSv2 ocean component was derived from the last 20 years of a 30-year simulation starting from CFSR state on November 1, 1980. The monthly anomalies of all variables from the ocean analyses are then calculated with respect to their own climatologies and superimposed on the CFSv2 monthly climatological states. The fields in March and April are averaged to represent the oceanic states at the start of April. Initializing the hindcasts using the monthly oceanic analyses is different from the operational practice of using an instantaneous analysis from the ocean data assimilation system. A set of test runs in Zhu et al. (2012b) showed that using the monthly fields as OIC has little impact on the deterministic forecasting skill (see Fig. S1 of the auxiliary material in Zhu et al. (2012b)).