A novel quintile multi-model ensemble approach for improving future extreme precipitation projections using XGBoost

Abstract A novel quintile multi-model ensemble (QMME) framework is introduced to improve future precipitation projections from MMEs. The QMME method divides daily precipitation into five quintiles, thereby capturing not only extreme rainfall but also the full spectrum of precipitation intensities. This approach surpasses conventional ensembles, such as the equal MME (EMME) and the weighted MME (WMME), in accurately reflecting the distinct characteristics of individual climate models. To develop the QMME, this study first applies empirical quantile mapping to correct biases in daily precipitation outputs from 14 coupled model intercomparison project 6 (CMIP6) general circulation models (GCMs). Historical observations (1980–2014) from 61-gauge stations in South Korea are then used to evaluate GCM performance within each quintile. The QMME framework integrates each GCM’s historical performance and future uncertainty into a quintile-specific weights and combines them with XGBoost to generate enhanced precipitation time series. Four shared socioeconomic pathway (SSP) scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are considered to assess a range of future scenarios. As a result, the QMME consistently outperforms both EMME and WMME in capturing both extreme precipitation events and moderate rainfall conditions. An evaluation using reliability ensemble averaging confirms that the QMME more effectively accounts for inter-model variability and reduces uncertainty, thus providing robust projections of future precipitation. This quintile-based methodology can be readily extended to other hydroclimatic variables and geographic regions, offering significant potential for improving climate impact assessments and guiding risk management in water resources planning.