Simulating sub-hourly rainfall data for current and future periods using two statistical disaggregation models: case studies from Germany and South Korea

Abstract

The simulation of fast-reacting hydrological systems often requires sub-hourly precipitation data to develop appropriate climate adaptation strategies and tools, i.e. upgrading drainage systems and reducing flood risks. However, these sub-hourly data are typically not provided by measurements and atmospheric models, and many statistical disaggregation tools are applicable only up to an hourly resolution. Here, two different models for the disaggregation of precipitation data from a daily to sub-hourly scale are presented. The first one is a conditional disaggregation model based on first-order Markov chains and copulas (WayDown) that keeps the input daily precipitation sums consistent within disaggregated time series. The second one is an unconditional rain generation model based on a double Poisson process (LetItRain) that does not reproduce the input daily values but rather generates time series with consistent rainfall statistics. Both approaches aim to reproduce observed precipitation statistics over different timescales. The developed models were validated using 10 min radar data representing 10 climate stations in Germany and South Korea; thus, they cover various climate zones and precipitation systems. Various statistics were compared, including the mean, variance, autocorrelation, transition probabilities, and proportion of wet period. Additionally, extremes were examined, including the frequencies of different thresholds, extreme quantiles, and annual maxima. To account for the model uncertainties, 1000-year-equivalent ensembles were generated by both models for each study site. While both models successfully reproduced the observed statistics, WayDown was better (than LetItRain) at reproducing the ensemble median, showing strength with respect to precisely refining the coarse input data. In contrast, LetItRain produced rainfall with a greater ensemble variability, thereby capturing a variety of scenarios that may happen in reality. Both methods reproduced extremes in a similar manner: overestimation until a certain threshold of rainfall and underestimation thereafter. Finally, the models were applied to climate projection data. The change factors for various statistics and extremes were computed and compared between historical (radar) information and the climate projections at a daily and 10 min scale. Both methods showed similar results for the respective stations and Representative Concentration Pathway (RCP) scenarios. Several consistent trends, jointly confirmed by disaggregated and daily data, were found for the mean, variance, autocorrelation, and proportion of wet periods. Further, they presented similar behaviour with respect to annual maxima for the majority of the stations for both RCP scenarios in comparison to the daily scale (i.e. a similar systematic underestimation).