Nationwide Radar-Based Precipitation Nowcasting—A Localization Filtering Approach and its Application for Germany

One approach to improve advection methods is the short-term ensemble prediction system (STEPS). STEPS decomposes precipitation fields into different spatial scales and filters those having a short lifetime. The latter is achieved by using an auto-regressive (AR) model that considers a sequence of observations. However, such a model tends to smooth nowcasting fields especially in small but convective precipitation areas and at longer lead-times. With focus on the deterministic configuration of STEPS, i.e., the spectral prognosis model (SPROG), this article 1) extends the STEPS approach by estimating spatially localized parameters of the AR process, 2) conducts a sensitivity analysis of the SPROG model to the order of the AR process, the spatial decomposition levels, and post-processing, and 3) analyzes the forecast skill of the extended STEPS. For such purpose, the performance of the localized AR model was demonstrated and evaluated at several precipitation thresholds and window sizes using a varied set of precipitation events collected by the radar network of the German Weather Service. The statistical results exhibited an improved performance of the localized AR model over SPROG when both are evaluated at precipitation thresholds and window sizes larger than 0.1 mm h
$^{-1}$
and 1 km, respectively, and for lead-times up to 2 h. The analysis suggested a first-order AR process, six cascade levels, and a mean adjustment post-processing procedure. Our results show a key role of the localization aspect when generating nationwide forecasts in scenarios that include large precipitation areas which are non-uniformly distributed having isolated convective features.