A spatial and seasonal climatology of extreme precipitation return-levels: A case study

Abstract

A spatial and seasonal modeling approach for precipitation extremes is introduced and exemplified for the Berlin-Brandenburg region in Germany. Monthly maxima of daily precipitation sums are described with a generalized extreme value distribution (GEV) with spatially and seasonally varying parameters. This allows for a return-level prediction also at ungauged sites. The seasonality is captured with harmonic functions, spatial variations are modeled with Legendre polynomials for longitude, latitude and altitude. Interactions between season and space allow for a spatially varying seasonal cycle. Orders of the harmonic and Legendre series are determined using a step-wise forward regression approach with the Bayesian Information Criterion (BIC) as model selection criterion. The longest 80 series are used to verify the approach in a cross-validation experiment based on the Quantile Skill Score (QSS). The model presented describes the observations at all these stations more accurately than a GEV applied to each month and location separately. These improvements are due to the assumption of smoothly varying GEV parameters in time and space; information from neighboring observations in time and space are used to obtain parameters at a given location. Apart from robustness, this approach allows also a seasonally and spatially varying shape parameter and results are found to be more accurate.