An Operational Statistical Downscaling Prediction Model Of The Winter Monthly Temperature Over China Based On A Multi-Model Ensemble

The interannual variation of East Asian winter surface air temperature (SAT) exhibits considerable differences between individual months. However, the prediction skill of dynamical models is rather low; therefore, a statistical downscaling (SD) model was employed to improve the skill for 36 winters during 1983/84-2018/19 in a cross-validated method and for 6 winters during 2013/14-2018/19 in an independent verification. The first two observed empirical orthogonal function (EOF) modes of monthly mean SAT variability explain about 60%-70% of the total variance. Correlation analysis of EOF principal components (PCs) is applied to explore the predictability from key lower boundary anomalies, i.e., Arctic sea ice and sea surface temperature (SST), for SD models. Two previous or simultaneous predictors over distinct areas are identified by the PC correlation analysis for individual monthly China SAT. It is assumed that the simultaneous SST predictors from the multi-model ensemble (MME) of Climate Forecast System version 2 (CFSv2) and BCC_CSM1.1 m involved in the SD model are perfectly predicted, and the potential attainable predictability can be obtained to qualify the highest prediction skills of the SD model. During 1983/84-2018/19, the area-averaged temporal correlation coefficient (TCC) skill of the MME model is 0.10, 0.26 and 0.30 (0.27 at the 90% confidence level) for December, January and February, respectively. The significantly high skill of the cross-validated SD forecast mainly covers extensive domains in central China, and the area-averaged TCC skill is improved compared with the original MME prediction. The corresponding area-averaged attainable TCC skills are 0.26, 0.46 and 0.50 in the SD model, with reference to significant TCC regions in most parts of China. The pattern correlation coefficient skill of the original MME model is much lower than that of the SD model prediction. Independent SD forecasting for a recent 6-year period further reveals that the winter monthly China SAT is highly predictable by the SD model. It should be noted that the China SAT in December is less predictable than that in January and February by the dynamical and SD models, which may be due to the complex predictability source of December SAT. During January and February, the El Nino-Southern Oscillation signal and Indian Ocean dipole mode can be considered as one of the prominent sources of predictability leading to a better prediction skill.