Predictability and Prediction of Early- and Peak-summer East Asian rainfall

East Asian summer monsoon (EASM) rainfall has a profound influence on the lives of billions of people. The seasonal prediction of the EASM rainfall, however, has long been an outstanding challenge in climate science. Traditional seasonal forecast of EASM deals with JJA mean rainfall anomalies, which may not be the best strategy because the EASM rainy season is typically from May to August and pronounced differences exist between early summer (May-June, MJ) and peak summer (July-August, JA): both climatological mean states and the principal modes of interannual variability exhibit distinct spatial and temporal structures. The present study explores the sources and limit of the predictability of the early and peak summer rainfall over the East Asian (EA) region. Since the climate models' seasonal forecasts have rather limited skills, it is important to find the causes of the low skills, to improve seasonal prediction, and to better estimate the predictability of EASM rainfall. We address this issue by applying predictable mode analysis method. Four empirical modes of variability for peak summer rainfall are identified: (a) an equatorial western Pacific-EA teleconnection mode, (b) a western Pacific subtropical high-dipole feedback mode, (c) a central Pacific-ENSO mode, and (d) a Eurasian wave train mode. These modes are named according to the major sources of predictability. Based on the understanding of predictability sources for each mode, a suite of physical-empirical (P-E) models is established to predict the four leading principal components (PCs). All four modes can be predicted with significant cross-validated correlation skills(0.59-0.65). Using the predicted PCs and the corresponding observed spatial patterns, a 35-year cross-validated hindcast over the EA yields a domain-averaged TCC skill is 0.37, which is higher than the MME hind cast skill (0.13). The estimated potential attainable pattern correlation coefficient skill averaged over the entire domain is around 0.61, suggesting that the current models' prediction may have large rooms to improve.