Effects of Spatial Aggregation on the Accuracy of Statistically Downscaled Precipitation Predictions

Abstract

Abstract The effects of spatial aggregation on the skill of downscaled precipitation predictions obtained over an 8 × 8 km2 grid from circulation analogs for metropolitan France are explored. The Safran precipitation reanalysis and an analog approach are used to downscale the precipitation where the predictors are taken from the 40-yr ECMWF Re-Analysis (ERA-40). Prediction skill—characterized by the continuous ranked probability score (CRPS), its skill score, and its decomposition—is generally found to continuously increase with spatial aggregation. The increase is also greater when the spatial correlation of precipitation is lower. This effect is shown from an empirical experiment carried out with a fully uncorrelated dataset, generated from a space-shake experiment, where the precipitation time series of each grid cell is randomly assigned to another grid cell. The underlying mechanisms of this effect are further highlighted with synthetic predictions simulated using a stochastic spatiotemporal generator. It is shown 1) that the skill increase with spatial aggregation jointly results from the higher and lower values obtained for the resolution and uncertainty terms of the CRPS decomposition, respectively, and 2) that the lower spatial correlation of precipitation is beneficial for both terms. Results obtained for France suggest that the prediction skill indefinitely increases with aggregation. A last experiment is finally proposed to show that this is not expected to be always the case. A prediction skill optimum is, for instance, obtained when the mean areal precipitation is estimated over a region where local precipitations of different grid cells originate from different underlying meteorological processes.