Bayesian conditioning of a rainfall-runoff model for predicting flows in ungauged catchments and under land use changes Nataliya Bulygina, 1 Neil McIntyre, 1 and Howard Wheater 1,2 Received 25 February 2010 ; revised 5 November 2010 ; accepted 18 November 2010 ; published 2 February 2011. [1] A novel method is presented for conditioning rainfall-runoff models for ungauged catchment and land use impact applications. The method conditions the model on information from multiple regionalized response indices using a formal Bayesian approach. Two indices that hold information about soil type and land use effects are the base flow index from the Hydrology of Soil Type (HOST) classification and curve number from the U.S. Department of Agriculture’s Soil Conservation Service soil and land use classification. These indices are used to constrain a five-parameter probability distributed moisture model for subcatchments of the Wye (grazed grassland) and Severn (mainly afforested) catchments in the United Kingdom. The base flow index and curve number constrain only two of the five model parameters, indicating that ideally, other sources of information would be sought. Nevertheless, the procedure significantly reduces the prior uncertainty in runoff prediction and gives predictions close to those of the calibrated models. For the case study, the introduction of the curve number in addition to the base flow index has only a small effect on model performance and uncertainty; however, it allows a distinction between the effects of soil type and land management for the purpose of scenario analysis. The principal assumptions used in the method are the applicability of the curve number classification system and its mapping to UK soil types and the likelihood function used for Bayesian conditioning. Citation: Bulygina, N., N. McIntyre, and H. Wheater (2011), Bayesian conditioning of a rainfall-runoff model for predicting flows in ungauged catchments and under land use changes, Water Resour. Res., 47, W02503, doi:10.1029/2010WR009240. 1. Introduction [2] Although there has been a 40 year history of success- ful application of hydrological models to simulate rainfall- runoff processes in gauged catchments, several problems remain fundamental challenges. Two of these are the repre- sentation of flow in ungauged catchments and the represen- tation of nonstationarity in catchment response, in particular, the effects of rural land use and land manage- ment change. There has been extensive discussion of the role and limitations of physics-based models in addressing these problems [e.g., Beven and Binley, 1992; Wheater et al., 1993; Beven, 2000, 2001; Wheater, 2002]. While Jackson et al. [2008] provide a strategy for the use of physics-based models to represent the effects of land use and land management change based on detailed experimen- tal data, in general, the data support for such approaches is limited, and the application of such models to ungauged catchments is associated with high levels of uncertainty, reflecting uncertainty in the prior distribution of parameter values [e.g., Lukey et al., 2000]. In this paper we focus on the use of simpler, conceptual model structures, with more parsimonious parameterizations, and consider the potential for conditioning based on regional data. [3] In conceptual modeling, it is well known that catch- ment physical properties cannot be used directly as model parameters ; hence, an alternative strategy for parameter specification is required. The problem is made harder by the fact that the strategy often cannot rely on fitting the model to observed hydrological data since a common pur- pose of modeling is not just to emulate observed responses but also to predict responses at ungauged locations or under future land use or land management changes. Conse- quently, parameters are often estimated using statistical relationships between parameter values and physical prop- erties of the catchment, called parameter regionalization. This has been tackled using at least two different general approaches. The first links model parameters directly to physical catchment characteristics (e.g., catchment area, steepness, soil permeability, and geographical location [Lamb and Kay, 2004; Lee et al., 2006; McIntyre et al., 2005; Young, 2006]), and the second conditions parameters on flow response indices (e.g., mean annual discharge and daily discharge standard deviation) that have previously been regionalized [Bardossy, 2007; Yadav et al., 2007; Zhang et al., 2008]. While the former approach has been more common, an advantage of the latter is that a number of regional models linking flow indices to catchment prop- erties are available [e.g., Boorman et al., 1995; U.S. Department of Agriculture (USDA), 1986], hence avoiding, or at least reducing, the need to build new regional models. [4] Despite the considerable research into rainfall-runoff model regionalization, arguably, there are no satisfactory methods for modeling the effects of rural land use and land management. This is because there are few data on the 1 Department of Civil and Environmental Engineering, Imperial College London, London, UK. 2 School of Environment and Sustainability, University of Saskatche- wan, Saskatoon, Saskatchewan, Canada. Copyright 2011 by the American Geophysical Union. 0043-1397/11/2010WR009240 W02503 1 of 13 WATER RESOURCES RESEARCH, VOL. 47, W02503, doi :10.1029/2010WR009240, 2011