Search

Study region: The Guangdong-Hong Kong-Macao Greater Bay Area (GBA), China. Study focus: Using hourly rain gauge data and CMORPH data, we use the duration-dependent generalized extreme value (d-GEV) model and the scaling invariant GEV model inferred by the Bayesian hierarchical model to derive the intensity-duration-frequency IDF characteristics of extreme precipitation in GBA and adjust their uncertainties. New hydrological insights for the region: The GEV location and scale parameters of IDF curves in GBA show similar spatial distribution and the higher-resolution CMORPH can capture more local details than rain gauge data. Meanwhile, compared with the rain gauge data, CMORPH produces significantly lower rainfall intensity of storms with short durations, which leads to large uncertainties of IDF curves derived from CMORPH for the short-duration rainfall. Additionally, the uncertainties of IDF curves can be substantially reduced by using the scaling invariant model that was inferred by the Bayesian hierarchical model, compared with the ordinary d-GEV method. Therefore, the Bayesian inference is suggested to be adopted for regional estimation of IDF curves, especially for regions of limited sub-daily gauge data.

Given that long time series for temporally highly resolved precipitation observations are rarely available, it is necessary to pool information to obtain reliable estimates of the distribution of extreme precipitation, especially for short durations. In this study, we use a duration-dependent generalized extreme value distribution (d-GEV) with orthogonal polynomials of longitude and latitude as spatial covariates, allowing us to pool information between durations and stations. We determine the polynomial orders with step-wise forward regression and cross-validated likelihood as a model selection criterion. The Wupper River catchment in the West of Germany serves as a case study area. It allows us to estimate return level maps for arbitrary durations, as well as intensity-duration-frequency curves at any location—also ungauged—in the research area. The main focus of the study is evaluating the model performance in detail using the Quantile Skill Index, a measure derived from the popular Quantile Skill Score. We find that the d-GEV with spatial covariates is an improvement for the modeling of rare events. However, the model shows limitations concerning the modeling of short durations d ≤ 30 min. For ungauged sites, the model performs on average as good as a generalized extreme value distribution with parameters estimated individually at the gauged stations with observation time series of 30–35 years available. [ABSTRACT FROM AUTHOR]