A Statistical Downscaling Model for Forecasting Summer Rainfall in China from DEMETER Hindcast Datasets

Abstract

Abstract A new statistical downscaling (SD) scheme is proposed to predict summertime multisite rainfall measurements in China. The potential predictors are multiple large-scale variables from operational dynamical model output. A key step in this SD scheme is finding optimal predictors that have the closest and most stable relationship with rainfall at a given station. By doing so, the most robust signals from the large-scale circulation can be statistically projected onto local rainfall, which can significantly improve forecast skill in predicting the summer rainfall at the stations. This downscaling prediction is performed separately for each simulation with a leave-one-out cross-validation approach and an independent sample validation framework. The prediction skill scores exhibited at temporal correlation, anomaly correlation coefficient, and root-mean-square error consistently demonstrate that dynamical model prediction skill is significantly improved under the SD scheme, especially in the multimodel ensemble strategy. Therefore, this SD scheme has the potential to improve the operational skill when forecasting rainfall based on the coupled models.