Peak-summer East Asian rainfall predictability and prediction part II: extratropical East Asia

The part II of the present study focuses on northern East Asia (NEA: 26 degrees N-50 degrees N, 100 degrees-140 degrees E), exploring the source and limit of the predictability of the peak summer (July-August) rainfall. Prediction of NEA peak summer rainfall is extremely challenging because of the exposure of the NEA to midlatitude influence. By examining four coupled climate models' multi-model ensemble (MME) hindcast during 1979-2010, we found that the domain-averaged MME temporal correlation coefficient (TCC) skill is only 0.13. It is unclear whether the dynamical models' poor skills are due to limited predictability of the peak-summer NEA rainfall. In the present study we attempted to address this issue by applying predictable mode analysis method using 35-year observations (1979-2013). Four empirical orthogonal modes of variability and associated major potential sources of variability are identified: (a) an equatorial western Pacific (EWP)-NEA teleconnection driven by EWP sea surface temperature (SST) anomalies, (b) a western Pacific subtropical high and Indo-Pacific dipole SST feedback mode, (c) a central Pacific-El Nino-Southern Oscillation mode, and (d) a Eurasian wave train pattern. Physically meaningful predictors for each principal component (PC) were selected based on analysis of the lead-lag correlations with the persistent and tendency fields of SST and sea-level pressure from March to June. A suite of physical-empirical (P-E) models is established to predict the four leading PCs. The peak summer rainfall anomaly pattern is then objectively predicted by using the predicted PCs and the corresponding observed spatial patterns. A 35-year cross-validated hindcast over the NEA yields a domain-averaged TCC skill of 0.36, which is significantly higher than the MME dynamical hindcast (0.13). The estimated maximum potential attainable TCC skill averaged over the entire domain is around 0.61, suggesting that the current dynamical prediction models may have large rooms to improve. Limitations and future work are also discussed.