Estimating pros and cons of statistical downscaling based on EQM bias adjustment as a complementary method to...

The increasing availability of coarse-scale climate simulations and the need for ready-to-use high-resolution variables drive the climate community to the challenge of reducing computational resources and time for downscaling purposes. To this end, statistical downscaling is gaining interest as a potential strategy for integrating high-resolution climate information obtained through dynamical downscaling over limited years, providing a clear understanding of the gains and losses in combining dynamical and statistical downscaling. In this regard, several questions can be raised: (i) what is the performance of statistical downscaling, assuming dynamical downscaling as a reference over a shared time window; (ii) how much the performance of statistical downscaling is affected by changes in the number of years available for training; (iii) how does the climate normal considered for the training affect the predictions. This study addresses these issues by applying a statistical downscaling procedure based on the empirical quantile mapping bias adjustment, obtaining finer-resolution climate variables. This procedure was adopted in order to downscale temperature and precipitation from ERA5 climate reanalysis, having as reference both for training and validation, the respective variables obtained through the dynamical downscaling of ERA5 over Italy for about 30 years. The availability of such a long simulation allows us to define several long time windows, used to calibrate the statistical relationships and evaluate the performance of statistical downscaling versus dynamical downscaling over a shared blind prediction period, taking advantage of a set of spatial and temporal metrics. The study shows that (i) the statistical downscaling successfully represents mean values and extremes of temperature and precipitation; (ii) its performance remains satisfactory regardless of the number of years used as training; (iii) the shorter is the time window considered for the training, the higher is the sensitivity to changes in the time interval due to the inter-annual variability. Nevertheless, the performance deviations are somehow not so remarkable.