Identify optimal predictors of statistical downscaling of summer daily precipitation in China from three-dimensional large-scale variables

Statistical downscaling (SD) of daily precipitation is a challenging task, and the identification of predictors is crucial for constructing SD models. This study focuses on identifying SD predictors for summer (June-September) daily precipitation in China. Six large-scale variables (LSVs) in ERA-Interim reanalysis were used to select predictors for 177 sites. For each site, the predictor identification was conducted by searching the grid box having the best correlation to precipitation in a three-dimensional way: across different grid boxes and multiple pressure levels. The result indicates that correlations are often sensitive to the pressure levels. Adjacent sites share similar spatial patterns of correlations, indicating regionally different physical relations between LSVs and precipitation. The predictor selection reasonably reflects the regional circulations related to precipitation. Twelve candidate predictors were used to train generalized linear models by least absolute shrinkage and selection operator (LASSO) algorithm. The validation indicates the models have generally high performance, and also shows relatively poor performance for the sites in North China, Northwest China, and Yunnan when compared to that in the east of China. The downscaled outputs can roughly reflect the annual variations of summer total precipitation and rainy days. Two experiments on the stationarity assumption of the models under different climate conditions were conducted, indicating that no areas/sites were found significantly violated the stationarity assumption. This study presents guidance on how to select suitable predictors for downscaling daily precipitation in different areas of China.