Hindcast experiments of JMA-CGCM02
A set of 119 Hindcast experiments of 17-month prediction from January 1985 to February 2000 was performed. Figure.1 shows all of the predicted sea surface temperature (SST) deviations from 1961-1990 means for Region B (4ºN-4ºS, 150ºW-90ºW) since 1988 together with the observed one. Due to bad qualities of the initials forecasts starting from 1985 to 1987 are poor and thus omitted.
JMA-CGCM02 properly predicts the variations of Region B SST. In particular, the occurrence of the 1997/98 El Niño event is well predicted with more than six months in advance though the predicted amplitude is a little bit smaller.
Figure.1 SST deviations from 1961-1990 mean for Region B from January 1988 to February 2000. Thick red line indicates observation and thin lines 17-month forecasts of MA-CGCM02.
Figure.2 shows anomaly correlation coefficient and root mean square error (RMSE) of Region B SST of hindcast of the new model (88 experiments on January 1989 - January 2000) and the former model (116 experiments on February 1989 - November 2000). Anomaly correlation coefficient for the new model is improved by about 0.05-0.3 compared with the former model. As of RMSE, the skill of the new model exceeds that of persistent forecast after two-three months lead time, while that of the former model does after five-six months. Performance of the new model is improved, especially within shorter lead time until seven to eight months.
Figure.2 Anomaly correlation coefficient (top) and RMSE (bottom) of the new model and the former model
Red line indicates forecasts of the new model and blue indicates those of the former model. Green and red line indicate persistent and climatological forecasts, respectively.
MOS forecast with JMA-CGCM02
The verification of MOS product should be done using independent samples. Due to lack of independent samples, however, enhancement of the forecast skill by MOS is verified using the dependent samples. Figure 3 shows anomaly correlation coefficient and RMSE (minus bias) of MOS forecast and model forecast of the new model (96 samples; top figure) and the former model (132 samples; bottom figure) since 1988. MOS correction works more effectively in the new model than in the former model because of the enhanced model performance.
Anomaly correlation coefficient of the new MOS forecast is larger than 0.8 until nine to ten months lead time, while that of the former MOS forecast until seven to eight months. Thus, it can be expected that new MOS forecast improves predictability even in independent forecasts.
New model
Former model
Figure 3 Anomaly correlation coefficient and RMSE of MOS forecast and model forecast of the new model (top) and the former model (bottom)
Red and blue line indicate anomaly correlation coefficient of MOS products and model outputs, respectively. Pink and light blue line indicate RMSE of MOS products and model outputs, respectively.