An elegant solution to characterise the residual errors in radar precipitation estimates is to generate an ensemble of precipitation fields. The paper proposes a radar ensemble generator designed for usage in the Alps using LU decomposition (REAL), and presents first results from a real-time implementation coupling the radar ensemble with a semi-distributed rainfall-runoff model for flash flood modelling in a steep Alpine catchment. Each member of the radar ensemble is a possible realisation of the unknown true precipitation field given the observed radar field and knowledge of the space-time error structure of radar precipitation estimates. Feeding the alternative realisations into a hydrological model yields a distribution of response values, the spread of which represents the sensitivity of runoff to uncertainties in the input radar precipitation field.
The presented ensemble generator is based on singular value decomposition of the error covariance matrix, stochastic simulation using the LU decomposition algorithm, and autoregressive filtering. It allows full representation of spatial dependence of the mean and covariances of radar errors. This is of particular importance in a mountainous region with large uncertainty in radar precipitation estimates and strong dependence of error structure on location. The real-time implementation of the radar ensemble generator coupled with a semi-distributed hydrological model in the framework of the forecast demonstration project MAP D-PHASE is one of the first experiments of this type worldwide, and is a fully novel contribution to this evolving area of applied research.