Combining runoff generating mechanisms and the Metastatistical Extreme Value approach to predict extreme floods in catchments with strong discontinuities in the flood frequency curve

River floods are the most common natural hazards worldwide and accurate flood estimation is crucial for reducing flood risk. Traditional flood frequency analysis relies on the assumption of homogeneity of the analysed floods. However, floods arise from multiple generating mechanisms, such as rainfall on wet and dry soils, rain-on-snow and snow-melt events. Streamflow records may therefore comprise mixtures of events. Ignoring this may cause significant errors in the estimation of flood frequency. The problem is particularly evident in catchments with a discontinuity in the flood frequency distribution, where the rarest floods are significantly larger than the rest of the events in the record. These situations cannot be represented by traditional frequency analyses. Extreme floods may thus occur unexpectedly and produce disproportionate losses and casualties. Here, we propose a practical method to handle the problem of flood frequency estimation in catchments with strong discontinuities in the flood frequency curves.

In this work, we focus on rivers among 160 case studies in Germany which show a marked discontinuity in the empirical flood frequency distribution and we use the simplified Metastatistical Extreme Value (SMEV) approach to separately include floods with different generating mechanisms in the estimation of the flood frequency distribution. We extract all the independent ordinary events from daily streamflow records and organize them into two groups according to the key runoff generation processes (rain-on-dry, mixture of rain-on-wet and snowmelt processes). We fit a statistical distribution (either power law or log normal based on the statistical properties of the ordinary events) to each group. Then, we use SMEV to calculate the emerging frequency distribution.

Our results show that the proposed approach improves the estimation of the magnitude of floods with long return periods. Considering the mixture of generating processes allows to reproduce the observed discontinuities in the flood frequency curves. Comparison with the standard Generalized Extreme Value distribution shows that the proposed method reduces the estimation bias, especially for large quantiles.

This study summarizes the results of the DFG-funded project "Propensity of rivers to extreme floods: climate-landscape controls and early detection - PREDICTED" (Deutsche Forschungsgemeinschaft - German Research Foundation, Project Number 421396820).