Regionalisation of the parameters of a hydrological model: Comparison of linear regression models with artificial neural nets

Abstract

In many practical applications of hydrological models, a case-specific parameter optimisation is not feasible due to lack of data. Model parameters may then be derived through regionalisation, i.e. by linking model parameters to physical catchment descriptors. Most studies reported in literature rely on linear regression techniques to establish the regionalisation scheme although it is well known that the underlying assumptions, e.g. the linearity of the relationship between physical catchment descriptors and model parameters, are often violated. Artificial neural networks (ANNs) are distribution-independent non-linear model structures and are therefore potentially useful tools for regionalising model parameters. This paper compared linear regression analysis and ANNs for regionalising the most sensitive parameters of the semi-distributed hydrological model SWAT (Soil and Water Assessment Tool) with an application in the Flemish part of the Scheldt river basin. The uncertainty on both regionalisation procedures was assessed with a non-parametric bootstrap method. ANNs delivered more accurate parameter estimates than linear regression equations if the non-linearities simulated by the ANNs have a physical meaning and if the physical catchment descriptors of the catchment under study lie within the range of the descriptor values of the sites used for the construction of the ANNs. For extrapolations outside this range, regression analysis yielded a better result. The uncertainty on the regionalised parameters was generally higher for ANNs than for regression equations, however, intervals for stream flow predictions were only slightly broader. Finally, the results of previous analyses were interpreted in function of the parameterisation of virtual land use scenarios. Because five out of seven model parameters depend on land use, it can be expected that regionalisation can enhance the estimation of non-crop related parameters of virtual land use scenarios.