Downscaling of DEMETER winter seasonal hindcasts over Northern Italy

Abstract

A novel method is applied in order to obtain winter predictions over Northern Italy using state-of-the-art multi-model seasonal ensemble hindcasts. The method consists of several stages. In the first stage, the best predictions are computed for a group of eight indices of large-scale circulation variability using the multi-model ensemble data set. The predictions are multiple linear regressions of single-model ensemble mean hindcasts produced within the European project DEMETER using six different coupled models. The regression is obtained using the method of the best linear unbiased estimate (BLUE). In the second stage, a standard statistical downscaling technique of the ‘perfect prog’ kind is applied in order to predict a group of 12 surface predictands starting from a group of predictors selected between the large-scale indices identified during the first stage. The selection of the predictands is carried out empirically, using those which lead to the best final prediction, while the regression coefficients are defined using observational data only, as in a ‘perfect prog’ downscaling technique. All steps of the prediction computation up to this point are performed in cross-validation mode. Finally, the full high-resolution surface winter predictions are reconstructed using an adequate selection of the forecasted predictands.

The predictions obtained have a much higher detail than the DEMETER direct model output predictions and, in parts of the domain, they are characterized by substantially significant skill. The improvement of the skill with respect to single-model ensembles is due to the use of the BLUE technique, while the statistical downscaling allows us to increase significantly the detail of the prediction. The study includes a discussion on the sensitivity of the results to both the period in years and the number of models used to produce the forecasts, and a comparison with the results obtained using a simple multi-model forecast in which all models are given the same weight.