Critical perspectives on the evaluation and optimization of complex numerical models of estuary hydrodynamics and sediment dynamics

Authors

  • J.R. French

    Corresponding author
    1. Coastal and Estuarine Research Unit, Department of Geography, University College London, Gower Street, London WC1E 6BT, UK
    • Coastal and Estuarine Research Unit, Department of Geography, University College London, Gower Street, London WC1E 6BT, UK
    Search for more papers by this author

Abstract

Numerical hydrodynamic and sediment transport models provide a means of extending inferences from direct observation and for advancing our understanding of estuarine processes. However, their parametric complexity invites questions concerning the extent to which model output can be assessed with respect to data. This paper examines the basis for evaluating the performance of complex hydrodynamic and sediment transport models, with reference to a case study of a muddy meso-tidal estuary. Sophisticated and computationally-intensive models should be evaluated using robust objective functions, but conventional measures of fit and model efficiency invoke restrictive assumptions about the nature of the errors. Furthermore, they offer little insight into causes of poor performance. Optimization of tidal hydrodynamic models can usefully combine conventional performance measures with harmonic analysis of modelled shallow water tidal constituents that are diagnostic of the interactions between tidal propagation, bathymetry and bottom friction. Models with similar efficiencies can thus be distinguished and likely sources of error pinpointed. Hydrodynamic models have a predictive power that is rooted in a more-or-less complete representation of the physical processes and boundary conditions that are well-constrained with respect to data. In contrast, fine sediment models rely on a less complete conceptualization of a broader set of processes and, crucially, have a parametric complexity that is unmatched by the quantity and quality of observational data. Their performance as measured by conventional objective functions is weaker and it is important to match the structural complexity of model errors with analyses that can localize the scales and times of poor performance. Wavelet analysis is potentially useful here as a means of identifying aspects of the model that need improvement. The context in which such models are deployed is also important. Used heuristically, what might otherwise be dismissed as weak models can still provide mechanistic support for empirically-derived inferences concerning specific aspects of system behaviour. Copyright © 2009 John Wiley & Sons, Ltd.

Ancillary