Large-scale hydrological modelling by using modified PUB recommendations: the India-HYPE case
Abstract. The scientific initiative Prediction in Ungauged Basins (PUB) (2003–2012 by the IAHS) put considerable effort into improving the reliability of hydrological models to predict flow response in ungauged rivers. PUB's collective experience advanced hydrologic science and defined guidelines to make predictions in catchments without observed runoff data. At present, there is a raised interest in applying catchment models to large domains and large data samples in a multi-basin manner, to explore emerging spatial patterns or learn from comparative hydrology. However, such modelling involves additional sources of uncertainties caused by the inconsistency between input data sets, i.e. particularly regional and global databases. This may lead to inaccurate model parameterisation and erroneous process understanding. In order to bridge the gap between the best practices for flow predictions in single catchments and multi-basins at the large scale, we present a further developed and slightly modified version of the recommended best practices for PUB by Takeuchi et al. (2013). By using examples from a recent HYPE (Hydrological Predictions for the Environment) hydrological model set-up across 6000 subbasins for the Indian subcontinent, named India-HYPE v1.0, we explore the PUB recommendations, identify challenges and recommend ways to overcome them. We describe the work process related to (a) errors and inconsistencies in global databases, unknown human impacts, and poor data quality; (b) robust approaches to identify model parameters using a stepwise calibration approach, remote sensing data, expert knowledge, and catchment similarities; and (c) evaluation based on flow signatures and performance metrics, using both multiple criteria and multiple variables, and independent gauges for "blind tests". The results show that despite the strong physiographical gradient over the subcontinent, a single model can describe the spatial variability in dominant hydrological processes at the catchment scale. In addition, spatial model deficiencies are used to identify potential improvements of the model concept. Eventually, through simultaneous calibration using numerous gauges, the median Kling–Gupta efficiency for river flow increased from 0.14 to 0.64. We finally demonstrate the potential of multi-basin modelling for comparative hydrology using PUB, by grouping the 6000 subbasins based on similarities in flow signatures to gain insights into the spatial patterns of flow generating processes at the large scale.