Geoscience Reference
In-Depth Information
Figure 4.13
Application of an SVM method to predict flood water levels in real time, with lead times of one
to six hours (after Yu et al., 2006, with kind permission of Elsevier).
normally included in ANN), and testing out the different possibilities for the kernel filters on the different
support vectors (similar to the choice of kernel in the hidden layer of the ANN method). As with the
ANN approach, there is no real mechanistic check on the resulting model, so that it is likely to perform
best in prediction for events that are within the range of the training data, and should be used with
care when extrapolating beyond that range to more extreme events since unreasonable predictions might
then occur.
An example result for a flood forecasting problem, with predictions for different lead times, is shown
in Figure 4.13 (taken from Yu
et al.
, 2006).
4.6.3 Classification and Regression Trees (CART)
The ANN and SVM methods of prediction are flexible methods of handling nonlinearities in the rainfall-
runoff relationships for a catchment but are essentially parametric methods, in that they require the
calibration of many different weighting coefficients given a training data set. There are some empirical
methods that can be considered non-parametric strategies for model development. These generally involve
classificatory algorithms in which the training data for the variables to be predicted are subdivided into
classes based on values of some input control variables. Those classes are expected to reflect the different
conditions that lead to different responses in the catchment. Thus, the most important input control
