Kriging‐variance‐based multi‐member ensembles of radar–rain‐gauge precipitation estimates: Application in Switzerland

Radar–rain‐gauge rainfall combination methods have been used extensively as meteorological applications. Such combinations are often achieved through geostatistical processes (kriging). In this work we use as a platform the system “CombiPrecip”, which is operational in MeteoSwiss and responsible for generating the radar–rain‐gauge combination product. The main motivation of this article is to produce multi‐member ensembles with a relevant spread constrained by the kriging variance. This involves exploring two questions: (a) We investigate to what extent the kriging variance of CombiPrecip is a satisfactory measure of uncertainty of the kriging expected value. We attempt to answer this question through a probabilistic verification of a 7‐year dataset against independent rain‐gauge measurements. This verification suggests that the probabilistic CombiPrecip output has skill, which remains satisfactory even for high intensities of precipitation. (b) We present an algorithm that combines the kriging expected value and variance of the CombiPrecip output with spatially autocorrelated noise images to produce ensembles of realistic members. We investigate the extent to which the texture and autocorrelation features of the members relate to that of the expected‐value CombiPrecip image, using radially averaged power spectra and semi‐variograms and we find a good agreement. Ensembles provided by this method may be useful as initial conditions for precipitation nowcasting systems. The proposed ensemble‐generation technique is not limited to geostatistics‐based applications but can be easily generalized to any methods that produce probabilistic outcomes.