Bias correction of RCM outputs using mixture distributions under multiple extreme weather influences

Abstract

The frequency and magnitude of water-related disasters such as floods and landslides have intensified due to climate change, especially over East Asia, including the South Korea region. In this region, extreme precipitation events originate from multiple sources, such as tropical cyclones (i.e., typhoons) and frontal synoptic systems. Climate scenarios generated by global climate models (GCMs) are employed to assess the future variations of extreme precipitation. Precipitation outputs from GCM scenarios must be localized via dynamic downscaling through regional climate models (RCMs). Bias correction is required to eliminate the biases between the RCM outputs and local observations. Quantile mapping, in which RCM output values are mapped by quantiles onto historical observed data of all precipitation except zero values by fitting a probabilistic distribution to each dataset, has been a popular technique for bias correction. In the current study, we tested several probabilistic distribution models. Additionally, we tested several mixture probabilistic distributions, combinations of traditionally employed distributions, because extreme precipitation events over South Korea can develop from multiple weather systems. We also tested traditionally employed distributions, such as exponential, gamma, and GEV distributions for precipitation values except zero values. Their performances were evaluated with various statistics, especially for extreme events, because the bias-corrected data should be used for the assessment of future variations of extreme precipitation. The results indicate that the tested mixture distributions are superior to traditional non-mixture distributions. The gamma-Gumbel mixture distribution showed the best performance in reproducing the statistical characteristics of especially extreme precipitation in a way that the majority of non-severe precipitation events are fitted to the gamma distribution, whose tail is light, and the extreme events are fitted to the Gumbel distribution. The future variations of extreme precipitation from climate scenarios such as RCP 4.5 and RCP 8.5 showed clear differences between probabilistic distribution models, indicating that the selection of an appropriate distribution is critical in the reasonable assessment of future extreme precipitation.