Estimation of Precipitation by Kriging in the EOF Space of theSea Level Pressure Field

Abstract

The term downscaling denotes a procedure in which local climatic information is derived from large-scale climate parameters. In this paper, the possibility of using as downscaling procedure a geostatistical interpolation technique known as kriging is explored. The authors present an example of the method by trying to reconstruct monthly winter precipitation in the Iberian Peninsula from the North Atlantic sea level pressure (SLP) field in wintertime (December–February). The main idea consists in reducing the spatial dimension of the large-scale SLP field by means of empirical orthogonal function (EOF) analysis. Each observed SLP field is represented by a point in this low-dimensional space and this point is associated with the simultaneously observed rainfall. New values of the SLP field, for instance, those simulated by a general circulation model with modified greenhouse gas concentrations, can be represented by a new point in the EOF space. The rainfall amount to be associated to this point is estimated by kriging interpolation in the EOF space. The results obtained by this geostatistical approach are compared to the ones obtained by a simpler analog method by trying to reconstruct the observed rainfall from the SLP field in an independent period. It has been found that, generally, kriging and the analog method reproduce realistically the long-term mean, that kriging is somewhat better than the analog method in reproducing the rainfall evolution, but that, contrary to the analog method, it underestimates the variance because of the well-known smoothing effect. It is argued that there exists an intrinsic incompatibility between the estimation of the mean and replication of the variability. Finally, both methods have been also applied to daily winter rainfall. The methods are also validated by downscaling winter precipitation from SLP. It is concluded that kriging yields a better estimation of daily rainfall than the analog method, but the latter better reproduces the probability distribution of rainfall amounts and of the length of dry periods.