River flow modelling using fuzzy decision trees

A modern real time flood forecasting system requires its mathematical model(s) to handle highly complex rainfall runoff processes. Uncertainty in real time flood forecasting will involve a variety of components such as measurement noise from telemetry systems, inadequacy of the models, insufficiency of catchment conditions, etc. Probabilistic forecasting is becoming more and more important in this field. This article describes a novel attempt to use a Fuzzy Logic approach for river flow modelling based on fuzzy decision trees. These trees are learnt from data using the MA-ID3 algorithm. This is an extension of Quinlan's ID3 and is based on mass assignments. MA-ID3 allows for the incorporation of fuzzy attribute and class values into decision trees aiding generalisation and providing a framework for representing linguistic rules. The article showed that with only five fuzzy labels, the FDT model performed reasonably well and a comparison with a Neural Network model (Back Propagation) was carried out. Furthermore, the FDT model indicated that the rainfall values of four or five days before the prediction time are regarded as more informative to the prediction than the more recent ones. Although its performance is not as good as the neural network model in the test case, its glass box nature could provide some useful insight about the hydrological processes.