Potential Skill Map of Predictors Applied to the Seasonal Forecast of Summer Rainfall in China

Abstract

Anomalous summer rainfall in China is affected by many factors, whose complex interaction restricts the predictability of Chinese summer rainfall (CSR). The predicting skill of the state-of-the-art dynamic models on the CSR is still limited, leaving challenges in developing objective statistical predicting methods. A method for searching potential predicting skill of predictors (i.e., potential skill map, PSM) is proposed, which can be used to select predictors automatically based on the PSM, and a new automatic statistical prediction model of the CSR is established.Compared with traditional linear correlation analysis, the PSM using the cross-validation concept not only reflects the potential predicting skills of predictors on predictands, but is free from effects of extreme events. It is completely based on real-time statistical predicting procedure, which aims to find sufficient conditions for predictands in logical. The PSM is an important supplement to the traditional correlation coefficient map. They work together to provide potential predictors with necessary and sufficient conditions. The predictor automatic selector takes advantage of the idea of ensemble forecasting. It selects predictors with the most significant potential forecasting skill from the PSM, and then generates final forecast products by averaging a large number of predicting members. The year-by-year automatic selection of the predicators is thus realized. This solution doesn't rely on subjective experiences of foreasters, and also provides a new way to further investigate the predictability of the interannual variability of the East Asian summer monsoon. This new automatic statistical prediction model of the CSR based on the PSM and the predictor automatic selector shows a high reforecast skill for the CSR. In the 21-year reforecasting experiment from 1999 to 2019, predictors in the previous autumn and winter seasons are used to predict the CSR. Results show an average symbol agreement rate of 60% and the mean anomaly correlation coefficient of 0.436 between the reforecast and the observed CSR. As to the predicting skill (PS) score in the National Climate Center, the reforecast CSR reaches 71.00 in average. After variance correcting, the PS score further increases to 82.10, which is much higher than predicting skills of current dynamical models. It is noteworthy that the reforecast experiment in the present uses the first 12 multiple regression coefficients and EOF modes of the CSM, of which the first 4 multiple regression coefficients and EOF modes play a dominant role in the overall distribution of the CSM. By contrast, higher-order modes could further improve the reforecast skill by increasing the diversity of the reforecasting CSM, which represent their potential physical implications.