Towards a consistent extreme value statistic across heterogeneous precipitation data

Statistically sound extreme‑value analysis is essential for modelling the return times of heavy rainfall events. To obtain statistically robust return times, long and homogeneous time series are required. Even with such series, the estimated return times are only valid for the data on which the statistics are derived. This makes it difficult to compare return times from different data sources and to classify the return times of forecasts, for which usually no sufficiently long time series exist. Nevertheless, reliable and comparable return times of heavy rainfall are crucial for design precipitation and for impact-oriented warnings.

Our aim is therefore to develop a method that enables a statistically consistent comparison of return levels from different datasets and, subsequently, to estimate return times of forecast precipitation values on a solid statistical basis.

We use three datasets: (1) spatially and temporally homogenized multiannual radar-based precipitation estimates (RADKLIM), (2) reanalysis data (COSMO‑REA6), and (3) the official rain gauge based design‑precipitation dataset for Germany (KOSTRA‑DWD2020). Because of its high quality and its status as the reference for hydraulic infrastructure in Germany, KOSTRA‑DWD2020 is regarded as the reference data. In a first step we compared the three datasets by analyzing their distributions, return times, and return levels. This comparison revealed a systematic, duration‑dependent underestimation of both RADKLIM and COSMO‑REA6 relative to KOSTRA‑DWD2020. Consequently, we are developing methods to correct this underestimation. Various approaches are being tested, such as the use of duration‑dependent correction factors or a duration‑spanning distribution‑mapping technique.

The resulting adjusted return times may be incorporated as additional information in effective warnings for heavy‑rainfall events. In the future, we plan to combine these return times with further impact‑oriented data to develop more comprehensive, impact‑oriented products.