A comparative assessment of cluster-based regionalization approaches using conceptual rainfall–runoff models

Accurate streamflow prediction in ungauged catchments remains a central challenge in hydrology. We investigated the utility of regionalization methods to explore how hydrologically informed clustering influences parameter transferability and performance of different runoff models applied over a national scale. This study evaluates the performance of conceptual hydrological models (GR4J, GR6J, HBV) using support vector regression (SVR), k-nearest neighbours and weighted average parameter (WAP) regionalization approaches across catchments in the CAMELS-GB dataset. The HBV model performed best in montane northern Britain where snow melt contributes to seasonal flows but otherwise struggled to capture the complexity of urbanised and agriculturally modified lowlands in southern and eastern Britain where more process-explicit and parameterised GR4J and GR6J models excelled in performance. Our results demonstrate that SVR provides a robust and flexible regionalization framework by effectively capturing nonlinear catchment–response relationships to outperform traditional similarity-based approaches across most hydrological regimes. Clustering also enhanced WAP regionalization for GR4J and GR6J, whereas HBV remained largely unaffected due to its higher parameter dimensionality and internal compensation. Our findings reveal that parameter transferability is strongly cluster-dependent and regionalization strategies must explicitly account for interactions between model structure, regionalization method, and dominant regional hydrological characteristics. The significance of the contribution lies in the generalizability of the approach to ungauged basins where robust flow predictions are required for water resource management, risk assessment and climate change adaptation studies worldwide.