Classification of river networks for prediction in ungauged basins
The majority of the world's river basins remain ungauged and, therefore, the triedand- tested empirical techniques for predicting floods and droughts cannot be applied. An alternative approach, which is currently receiving a great deal of attention from research hydrologists, is to develop continuous simulation models whose parameters pertain to physical or hydrological properties of the river basins. However, difficulties related to scale, heterogeneity and complexity of real river basins have made a priori estimation of such parameters impossible: their estimation has always required calibration using river flow data. Therefore, estimating hydrological model parameters in ungauged river basins is one of the greatest challenges currently facing research hydrologists. In this thesis research advances towards this goal have been made at three different levels. First, at a conceptual level, a novel method for classifying river basins according to their physical properties is proposed. It is specifically designed for transferring hydrological model parameters from gauged river basins, where calibration is possible, to ungauged river basins. This approach relies on recognising that river basins can be similar in parts of their hydrological cycle but not in others. Thus, basins go through three independent classifications, one relative to each of the major components of the land phase hydrological cycle: interaction of soil water/vegetation and atmosphere; surface flow; and groundwater flow. This requires the ability to characterise the response of the components of the hydrological cycle independently, which leads to a second conceptual advance; rather than relying entirely on measured river flow data, from which it is difficult to separate out the effects of the three components, classification rules are devised on the basis of synthetic data produced by comprehensive, distributed, physically-based models. This thesis focuses on the surface flow component, applying the methodology to the identification of the best classifiers for surface flow through river networks. This required simulating river flow through a large number of Scottish river basins, which led to more practical research advances; all available commercial flow routing models were too cumbersome and required an impractical level of detail to be applied in such a large study. Therefore, a new flow routing modelling system was developed that extracts river network detail from digital databases and numerically solves a distributed flow routing model. Finally, on a detailed scientific level, significant insights have been made into the relationship between river network geomorphologic structure and stream flow response. In particular, it is shown that: a downstream hydraulic geometry relationship exists for Scottish rivers; although channel conveyance is a key factor in dictating network response, the features of the response hydro graph - namely the percentage attenuation of the flood peak and the lag in time to peak - scale linearly with both roughness and hydraulic geometry coefficients; much publicised invariant power law scaling rules for flood peaks in fact vary as a function of storm duration; statistical multivariate analysis of the simulated network flow responses demonstrated the low capacity of the network descriptors commonly used in regionalisation studies for characterising flow response. Four variables are shown to have significantly higher classifying power than the majority of the commonly used classifiers. Of these, two are entirely new to this thesis.