Parameter regionalization of large-scale distributed rainfall–runoff models using a conditional probability method

Abstract Given the evident impact of climate change, the frequency of severe flood events has increased worldwide. For various risk-reduction measures, covering all rivers in a country or regions including small-to-medium-sized rivers, flood risk assessment and real-time forecasting based on large-domain and high-resolution distributed rainfall–runoff models are fundamental. Due to limited observed records in such small-to-medium-sized rivers, the used distributed model must be robust and physically sound with the regionalized model parameters. Specifically, rather than optimizing parameters in many independent river basins, leading to a patched parameter distribution, regionalization should reflect the spatial distribution of hydrological signatures, such as soil and geology types. However, optimizing the parameters with existing methods incurs computational costs, posing difficulties in the parameter regionalization of large-domain and high-resolution distributed runoff models. To address this challenge, we propose a parameter regionalization method based on conditional probability. The key feature of this method is that the calibration phase calculation assumes spatially uniform parameter sets within the calibrating basins, significantly reducing computational costs. However, the resulting parameter sets are spatially distributed corresponding to the region’s pre-prepared soil or geological maps. It was achieved by introducing the Bayes’ theorem to estimate the conditional probability of the parameter set. The proposed method was applied to the distributed rainfall–runoff–inundation (RRI) model developed for Japan with a resolution of 150 m. The model performance in the validation phase, in which the performance was evaluated with 2723 flood events at 711 gauging stations, the median Nash–Sutcliffe efficiency (NSE) being 0.87, comparable or even improved to the performance in the calibration phase (NSE = 0.83) with 525 flood events at 75 dam reservoirs. Overall, the obtained nationwide high-resolution model is robust with good performance, even in ungauged basins. Furthermore, the proposed regionalization is a simple and useful way reflecting spatially distributed hydrologic signatures in the model parameters, and it can be utilized for any distributed rainfall–runoff model.