A regionalized framework for the Metastatistical Extreme Value Distribution applied to daily and sub-daily rainfall

The estimation of extreme rainfall based on short records is of considerable interest, above all in the context of rapidly changing rainfall regimes. Regionalization techniques, by trading space for time, allow us to partially overcome the lack of long observational records. The recently introduced Metastatistical Extreme Value Distribution (MEVD), a non-asymptotic extreme-value model, accounting for all observed rainfall events to infer the probability distribution of annual maxima, also contributed towards improving our ability to determine large quantiles based on short observational time series. Here we combine established regionalization techniques, aggregating data from multiple adjacent stations complying with set homogeneity criteria, with MEVD-based methodologies to explore how their joint use may further reduce the predictive uncertainty associated with the estimates of the probability of large events. In this work, we use precipitation data sets from a selection of worldwide regional station networks (Europe, USA, Middle East, and Asia) deployed in a wide range of elevations and different rainfall regimes. The temporal data resolution varies according to country ranging from sub-daily to daily scales. We analyze different event durations, between 5 minutes and 24 hours for the sub-daily scale, 1 day and 2 days for the daily one, and we implement a cross-validation procedure to evaluate predictive uncertainty. To evaluate possible improvements with respect to regionalization techniques based on traditional extreme value theory, such as the Generalized Extreme Value (GEV) distribution, we comparatively apply them and the proposed MEVD-based regionalization approach. The results show the benefits arising from the regionalization technique, which enhances the robustness of the models by increasing the consistency of the observed data population within the stations of the same cluster, particularly in the lowlands, where homogeneous regions can be more trivially identified. The proposed regionalization approach based on the metastatistic distribution brings a significant reduction of the estimation uncertainty for very high ratios between the forecasting return period value and the length of the calibration sample when compared to traditional methods.