A Proposed Rule-Discovery Scheme for Regionalisation of Rainfall-Runoff Characteristics in New South Wales, Australia

A project which sets out to explore alternative methods of regionalisation for prediction of streamflow in ungauged catchments is described. A commonly applied regionalisation methodology estimates model parameters for the ungauged basin from multiple linear regressions between parameters and catchment attributes, calibrated on gauged catchments. Its effectiveness is limited by the availability of sufficiently similar, well gauged and modelled catchments. In this paper a different approach is adopted, employing a data mining methodology to seek a useful set of rules in the catchment attribute space with parameter values as consequent. The outline of a hypothesis testing algorithm is given, along with a suggestion for a new discretisation method. The techniques are both based on the information theoretic concepts of cross entropy and mutual information.