Multiscale spatial recorrelation of RCM precipitation to produce unbiased climate change scenarios over large areas and small
Previously, Bárdossy and Pegram (2011) achieved downscaling of regional climate model (RCM) rainfall, dependent on circulation patterns (CPs), over 172 areas of the Rhine basin at the 25 km scale. Uneasy about the spatial statistics of the downscaled RCM rainfall, we calculated the spatial cross-correlation coefficients (CCCs) of daily rainfalls of the same set. We found that the CCCs of the RCM precipitations were significantly lower than those of the observations. CP-based downscaling led to an increase of the CCCs which still remained below the observed CCCs. This underestimation of spatial correlation, hence observed clustering, has potentially deleterious consequences for flood calculations over large areas based on RCM outputs, even after full CP-based bias elimination at the 25 km scale. In this paper we therefore describe two novel recorrelation methods designed to correct the CCCs of the RCM estimates back to those of the observed set before undertaking the final quantile-quantile transform. We use two methods of recorrelation: matrix methods and sequential regression. They both produced similar results and were successful in that they captured the observed CCCs almost exactly, coping with problems presented by the high proportion of dry days. In spite of the complete success of the recorrelation techniques (when comparing spatial correlations before and after treatment) the methodology does not solve the reconstitution problem fully: (1) extreme daily rainfall totals on large areas are not recaptured completely and (2) clustering behavior, as computed by entropy on nonoverlapping triple sites, confirms that the two-dimensional covariance dependence measure, although very effective, does not capture all of the clustering observed in natural rainfall.