Validating Hydrological Models In A Data Scarce Country – Getting The Right Results For The Right Reasons

This study explores some of the principle issues associated with quantifying surface  water and groundwater interactions and the practical application of models in a data scarce region such as South Africa. The linkages between the various interdependent components of the water cycle are not well understood, especially in those regions that suffer problems of data scarcity, and there remain  urgent  requirements  for  regional  water  resource  assessments.  Hydrology  (both  surface water and groundwater hydrology) is a difficult science; it aims to represent highly variable and non- stationary processes which occur in catchment systems, many of which are unable to be measured at the scales of interest. The conceptual representations of these processes are translated into mathematical form in a model. Different process interpretations, together with different mathematical representations, result in the development of diverse model structures. These structural uncertainties are difficult to resolve due to the lack of relevant data. Further uncertainty is introduced  when  parameterising  a  model,  as  the  more  complex  the  model,  the  greater  the possibility that many different parameter sets within the model structure might give equally acceptable results when compared with observations. Incomplete and often flawed input data are then used to drive the models and generate quantitative information. Approximate implementations (model structures and parameter sets), driven by approximate input data, will necessarily produce approximate results. Most model developers aim to represent reality as far as possible, and as our understanding of hydrological processes has improved, models have tended to become more complex. Beven (2002) highlighted the need for a better philosophy toward modelling than just a more explicit representation of reality and argues that the true level of uncertainty in model predictions  is  not  widely  appreciated.  Model  testing  has  limited  power  as  it  is  difficult  to differentiate  between  the  uncertainties  within  different  model  structures,  different  sets  of alternative parameter values and in the input data used to run a model. A number of South African case studies are used to examine the types of data typically available and explore the extent to which a model is able to be validated considering the difficulty in differentiating between the various sources of uncertainty. While it is difficult to separate input data, parameter and structural uncertainty, the study found that it should be possible to at least partly identify the uncertainty by a careful examination of the evidence for specific processes compared with the conceptual structure of a specific model. While the lack of appropriate data means there will always be considerable uncertainty surrounding model validation, it can be argued that improved process understanding in an environment can be used to validate model outcomes to a degree, by assessing whether a model is getting the right results for the right reasons.