On evaluation of ensemble precipitation forecasts with observation-based ensembles

Spatial interpolation of precipitation data is uncertain. How important is this uncertainty and how can it be considered in evaluation of high-resolution probabilistic precipitation forecasts? These questions are discussed by experimental evaluation of the COSMO consortium’s limitedarea ensemble prediction system COSMO-LEPS. The applied performance measure is the often used Brier skill score (BSS). The observational references in the evaluation are (a) analyzed rain gauge data by ordinary Kriging and (b) ensembles of interpolated rain gauge data by stochastic simulation. This permits the consideration of either a deterministic reference (the event is observed or not with 100% certainty) or a probabilistic reference that makes allowance for uncertainties in spatial averaging. The evaluation experiments show that the evaluation uncertainties are substantial even for the large area (41 300 km2) of Switzerland with a mean rain gauge distance as good as 7 km: the oneto three-day precipitation forecasts have skill decreasing with forecast lead time but the oneand two-day forecast performances differ not significantly.