Application of a Nearest Neighbour Method to a Conceptual Rainfall - Runoff Model

Current regionalization techniques for conceptual rain-fall runoff models employ linear/nonlinear regression functions. The model parameters are described as functions (assumed a priori) of catchment descriptors. As the hydrological system is highly non-linear and complex, no single parametric function can be a holistic or near-reality representation of the core underlying processes. In this paper, the Nearest Neighbor Method (NNM) is used to tackle the aforementioned problem. NNM is essentially a non-parametric approach. It is used to classify hydrological units based on their distance. The distance is not geo-physical but one in a transformed space. The classification is established by searching for a transformation matrix through optimization towards a globally optimal objective function. Optimization is carried out by a simulated annealing algorithm. Hydrological neighboring units in the transformed space are then classified using an extended nearest neighbor method. When the output of one hydrological unit is to be predicted, its so-called neighbors’ response is used by taking a regression or kriging over the group of ‘close’ neighbors. This approach is finally extended to identify the link between model parameters and catchment characteristics.